Plaintext
Article
Speech Engines
†
James Grimmelmann
Introduction ................................................................................ 869
I. How Search Engines Work ................................................... 876
II. The Conduit and Editor Theories ....................................... 879
A. The Conduit Theory .................................................. 880
B. The Editor Theory..................................................... 885
C. Three Points of Disagreement ................................. 889
III. The Advisor Theory ............................................................ 893
A. Search Results as Advice ......................................... 893
B. Active Listening ........................................................ 896
C. Access and Loyalty.................................................... 901
1. Access ................................................................... 901
2. Loyalty ................................................................. 903
3. Access and Loyalty Compared............................ 906
D. Limits of the Advisor Theory ................................... 906
IV. Search Bias Reconsidered .................................................. 911
A. Search Rankings Are Opinions About Relevance... 913
B. The First Amendment Requires Falsity and
† Professor of Law, University of Maryland. I presented earlier versions
of this work in faculty workshops at the University of Michigan, George Wash-
ington University, the University of Maryland, and Georgetown; to the Yale
Law School Information Society Project; and at the 2012 Intellectual Property
Scholars and 2013 Free Expression Scholars conferences. My thanks to the
participants, and to Aislinn Black, Jack Balkin, Yochai Benkler, Stuart Ben-
jamin, Shyamkrishna Balganesh, Bruce Boyden, Oren Bracha, Donald
Braman, Michael Carroll, Anne Chen, Danielle Citron, Richard Chused, Julie
Cohen, Laura DeNardis, Deven Desai, Michael Froomkin, Doni Gewirtzman,
Tarleton Gillespie, Eric Goldman, Joris van Hoboken, David Johnson, Eddan
Katz, Molly Land, Sarah Lawsky, Thomas Lee, Mark Lemley, Viva Moffat,
Paul Ohm, Frank Pasquale, Mark Patterson, Guy Pessach, Lisa Ramsey,
Chris Riley, Rebecca Roiphe, David Super, Peter Swire, Rebecca Tushnet, Fe-
lix Wu, and Steven Wu. Copyright © 2014 by James Grimmelmann. This Arti-
cle is available for reuse under the Creative Commons Attribution 3.0 United
States license, http://creativecommons.org/licenses/by/3.0/us. The required at-
tribution notice under the license must include the article’s full citation infor-
mation, e.g., “James Grimmelmann, Speech Engines, 98 MINN. L. REV. 868
(2014).”
868
2014] SPEECH ENGINES 869
Fault .......................................................................... 917
1. Milkovich.............................................................. 918
2. Some Objections .................................................. 919
C. Search Rankings Are False When They Are
Dishonest ................................................................... 922
1. The Conduit Theory: Objective Falsity.............. 923
2. The Editor Theory: No Falsity ........................... 924
3. The Advisor Theory: Subjective Falsity............. 926
D. Algorithms Are a Red Herring ................................. 932
E. Conclusion: The Federal Trade Commission Gets
It Mostly Right .......................................................... 934
V. Other Applications ............................................................... 936
A. Copyright ................................................................... 937
B. Privacy ....................................................................... 941
C. Defamation ................................................................ 943
D. Trademark ................................................................. 947
Conclusion................................................................................... 950
INTRODUCTION
To understand Google, there are worse places to look than
the New York Times editorial pages. Not because the Times has
some special insight into this search colossus, but rather pre-
cisely because it does not. In 2009 and 2013, the Times pub-
lished a pair of mirror-image op-eds, one each for and against
the company, presenting the toughest allegations against it and
the broadest defense of its actions. Each of them expresses
something like the conventional wisdom about Google. And in
the contrast between them can be seen something of why it is
1
so hard to know just what to do about search engines.
1. In a previous article, I presented a descriptive taxonomy of legal and
policy issues relating to search engines. See generally James Grimmelmann,
The Structure of Search Engine Law, 93 IOWA L. REV. 1 (2007). This Article
adds a normative framework for resolving those issues. Search engines are a
distinctive category of interest for reasons spelled out in more detail infra Part
III: their unique capacity to help numerous diverse users achieve their per-
sonal individual goals. Google will serve as the principal example, because of
the long shadow it casts on American law, culture, and technology, but most of
the discussion will be applicable more broadly to other search engines, such as
Bing and DuckDuckGo. The Article will use the term “search” as an abstract
noun to describe the social practice of using keyword-based search engines.
This social practice, even more than Google or other search engines, is the Ar-
ticle’s true subject; once it is rightly understood, the question of how to regu-
late Google becomes far clearer.
870 MINNESOTA LAW REVIEW [98:868
In December 2009, the Times ran Search, But You May Not
Find by Adam Raff, the co-founder of the United Kingdom-
based search engine Foundem, in which he accused Google of
2
slanting its search results to favor its own services. Raff wrote
that Foundem “was effectively ‘disappeared’ from the Internet”
3
when it was demoted in Google’s search results. He called on
the government to adopt a policy of “search neutrality” and pro-
4
tect websites like Foundem from Google’s dominance. His
charges anticipated the “search bias” issues at the heart of the
ambitious Federal Trade Commission (FTC) antitrust investi-
gation of Google: complaints that Google unfairly favored its
5
own maps and other specialized “vertical” content over others’.
But in January 2013, just after the FTC’s investigation had
fizzled out with a no-action letter on search bias, the Times ran
Is Google Like Gas or Like Steel? by Bruce Brown and Alan Da-
6
vidson. Google, they argued, was like the Associated Press:
7
protected by the First Amendment. In 1945, the Supreme
Court held that antitrust law would not “compel [the Associat-
ed Press] or its members to permit publication of anything
8
which their ‘reason’ tells them should not be published.” This
same standard should apply to Google, Brown and Davidson
argued, explaining that “search engines need to make choices
2. Adam Raff, Op-Ed., Search, But You May Not Find, N.Y. TIMES, Dec.
28, 2009, at A27.
3. Id. A search result is “demoted” when it moves from a prominent posi-
tion to a less visible one, e.g., from the top search result on the first page of
results to the sixth result on the twentieth page. Search rankings are de-
scribed in more detail infra Part I, and the difficulties of defining a baseline
against which to measure alleged demotion are discussed infra Part IV.
4. Id. (defining search neutrality as “the principle that search engines
should have no editorial policies other than that their results be comprehen-
sive, impartial and based solely on relevance”).
5. See Statement of the Fed. Trade Comm’n, In re Google Inc., No. 111-
0163 (Jan. 3, 2013). In addition to search bias, the FTC considered allegations
that Google was copying content without permission from websites, unfairly
making it difficult for advertisers to switch to competing search engines, and
misusing standard-essential patents, among other issues. See generally Deci-
sion and Order, In re Motorola Mobility, No. 12-0120 (Jan. 3, 2013). In the
end, it was this last prong that generated the strongest enforcement action.
This Article focuses on search bias because it raises the most truly fundamen-
tal issues about what we as a society want from search engines.
6. See Bruce D. Brown & Alan B. Davidson, Op-Ed., Is Google Like Gas
or Like Steel?, N.Y. TIMES, Jan. 4, 2013, at A17; Raff, supra note 2.
7. Id. (comparing the Associated Press’s concern “about a regulator in
every newsroom” with Google’s concern “about a regulator in every algo-
rithm”).
8. Associated Press v. United States, 326 U.S. 1, 20 n.18 (1945).
2014] SPEECH ENGINES 871
about what results are most relevant to a query, just as a news
editor must decide which stories deserve to be on the front
9
page.”
These op-eds endorse two diametrically opposed theories of
what a search engine is. To Raff, and to scholars like Jennifer
10 11
Chandler and Frank Pasquale, Google ought to be a passive
and neutral conduit, connecting users to websites and then
stepping out of the way. To Brown and Davidson, and to schol-
12 13
ars like Eric Goldman and Eugene Volokh, Google instead
ought to be an active and opinionated editor, sifting through
the Internet and using expert judgment to identify the im-
portant and the interesting. These two theories form the rhe-
torical backdrop to the ongoing legal battles over search.
The choice between “conduit” and “editor” has decisive im-
plications for how the law should deal with Google—and it is
more complicated than a simple “Google wins” or “Google los-
es.” On search bias claims like Foundem’s, the conduit theory is
a recipe for regulation, while the editor theory offers a First-
14
Amendment get-out-of-jail-free card. But when the issue is
defamation, the conduit theory holds Google harmless for the
sins of the websites it unknowingly connects users to, while the
editor theory calls down the vengeance of the heavens on
15
Google for its editorial decisions.
Indeed, not even Google itself can keep straight whether it
is an objective conduit or a subjective editor. In 2006, respond-
ing to a search-bias lawsuit from the children’s-information
website KinderStart, one of Google’s lawyers explained that
“Google is constantly evaluating Web sites for standards and
9. Brown & Davidson, supra note 6. Brown knows more than a little
about the press’s free speech rights: he is executive director of the Reporters
Committee for Freedom of the Press. Davidson—before a stint as Director of
Public Policy for the Americas for Google—worked on free expression issues
for the Center for Democracy and Technology. See Andrew McLaughlin,
Google Goes to Washington, GOOGLE OFFICIAL BLOG (Oct. 6, 2005), http://
googleblog.blogspot.com/2005/10/google-goes-to-washington.html.
10. Jennifer A. Chandler, A Right to Reach an Audience: An Approach to
Intermediary Bias on the Internet, 35 HOFSTRA L. REV. 1095 (2007).
11. E.g., Frank Pasquale, Rankings, Reductionism, and Responsibility, 54
CLEV. ST. L. REV. 115 (2006).
12. E.g., Eric Goldman, Search Engine Bias and the Demise of Search En-
gine Utopianism, 8 YALE J.L. & TECH. 188 (2006).
13. Eugene Volokh & Donald M. Falk, Google First Amendment Protection
for Search Engine Search Results, 8 J.L. ECON. & POL’Y 883 (2012).
14. See infra Part IV.
15. See infra Part V.C.
872 MINNESOTA LAW REVIEW [98:868
16
quality, which is entirely subjective.” But in 2012, Google
faced a defamation lawsuit from the former “First Lady” of
Germany, Bettina Wulff, who objected that typing [bettina
wulff] into Google produced autocomplete search suggestions
including [bettina wulff escort] and [bettina wulff
17
prostitute]. Google’s response: autocomplete suggestions
are “the algorithmic result of several objective factors, including
18
the popularity of search terms.” Google’s enemies are equally
opportunistic: Google should objectively present the web as it
16. Dawn Kawamoto, Suit over Poor Google Ranking May Go Forward,
CNET NEWS (June 30, 2006, 2:59 PM), http://news.cnet.com/Suit-over-poor
-Google-ranking-may-go-forward/2100-1025_3-6090239.html (quoting David
Kramer) (emphasis added). This is not an isolated statement; Google has made
it repeatedly to courts. See Kinderstart.com v. Google, Inc., No. C 06-2057 JF
(RS), 2007 WL 831806, at *20–21 (N.D. Cal. Mar. 16, 2007) (“Google itself
holds out PageRank as an opinion . . . .”); Langdon v. Google, Inc, 474 F. Supp.
2d 622, 630 (D. Del. 2007) (comparing Google to newspapers that retain the
freedom “to exercise subjective editorial discretion”); Kinderstart.com v.
Google, Inc., No. C 06-2057 JF (RS), 2006 WL 3246596, at *13–14 (N.D. Cal.
July 13, 2006) (“Google also contends . . . PageRank cannot be anything but
subjective.”); Search King, Inc. v. Google Tech., Inc., No. CIV-02-1457-M, 2003
WL 21464568, at *3–4 (W.D. Okla. May 27, 2003) (“Google argues that
PageRanks are subjective opinions . . . .”). Google’s Eric Schmidt has also
claimed, in Congressional testimony, “Search is subjective, and there’s no ‘cor-
rect’ set of search results.” The Power of Google: Serving Consumers or Threat-
ening Competition?: Hearing Before the Subcomm. on Antitrust, Competition
Pol’y and Consumer Rights of the S. Comm. on the Judiciary, 112th Cong. 238
(2011) (statement of Eric Schmidt, Executive Chairman, Google, Inc.) [herein-
after Schmidt Testimony].
17. Stefan Niggemeier, Autocompleting Bettina Wulff: Can a Google Func-
tion Be Libelous?, DER SPIEGEL (Sept. 20, 2012, 3:45 PM) (Paul Cohen, trans.),
http://www.spiegel.de/international/zeitgeist/google-autocomplete-former
-german-first-lady-defamation-case-a-856820.html. In this Article, I follow the
industry convention of placing search terms between brackets and setting
them in a fixed-width typeface.
18. Id. (emphasis added) (quoting unnamed Google spokesman) (internal
quotation marks omitted); see also An Explanation of Our Search Results,
GOOGLE, http://www.google.com/explanation.html (last visited Nov. 27, 2013)
(capturing Google’s current explanation: “The beliefs and preferences of those
who work at Google, as well as the opinions of the general public, do not de-
termine or impact our search results”). The initial version of this statement
was even more emphatic. An Explanation of Our Search Results, GOOGLE
(June 7, 2004), http://web.archive.org/web/20040607132019/http://www.google
.com/explanation.html (capturing Google’s message in 2004: “Our search re-
sults are generated completely objectively . . . . Because of our objective and
automated ranking system, Google cannot be influenced by these petitions [to
remove particular links or otherwise adjust search results]” (emphasis add-
ed)).
2014] SPEECH ENGINES 873
is, except when that would be bad, in which case Google should
19
subjectively step in to fix things.
Others have noted the tension between these two theories
20
of search. In First Amendment terms, the crucial difference is
the identity of the relevant speaker: the conduit theory focuses
on websites’ speech, the editor theory on search engines’
speech. Raff’s op-ed is an eloquent plea for Foundem’s right to
present its views to users free from Google’s interference;
Brown and Davidson’s op-ed is an equally eloquent plea for
Google’s right to present its own views to users free from the
government’s interference. Speech meets speech, with no obvi-
ous way to assign priority to one or the other. The Problem of
Google thus presents itself as an intractable opposition between
websites and search engines; it puts courts and regulators to a
stark and partisan choice between two incompatible characteri-
zations of search.
But there is another possibility. It takes two to tango, and
three to search. In addition to the website and the search en-
gine, there is also the user. For all their differences, the conduit
and editor theories have a common conception of search users:
as audiences, whose only job is to consume the speech of others.
On the conduit theory, users are eyeballs for websites; the
search engine’s job is to get out of the way and deliver to each
website the user traffic to which it is properly entitled. And on
the editor theory, users are eyeballs for search engines; a dis-
satisfied user’s best and only option is to change the channel
and try another search engine. Neither of these models fully
captures how people use search, because search responds to us-
ers’ interests in a way that other media do not.
Instead of passively consuming from a fixed menu set be-
fore them, search users actively seek out information. Each
query—[fayette monroe shooting] or [brining pheas-
ants for smoking] or [baby splotchy rash with white
19. For a good example of the kitchen-sink approach to attacking Google,
see generally SCOTT CLELAND, SEARCH & DESTROY: WHY YOU CAN’T TRUST
GOOGLE INC. (2011).
20. See, e.g., Frank Pasquale, Internet Nondiscrimination Principles:
Commercial Ethics for Carriers and Search Engines, 2008 U. CHI. LEGAL F.
263, 266 (questioning whether “dominant search engines” should have obliga-
tions or immunities); Danny Sullivan, KinderStart Becomes KinderStopped in
Ranking Lawsuit Against Google, SEARCH ENGINE WATCH (July 14, 2006),
http://searchenginewatch.com/article/2058241/KinderStart-Becomes
-KinderStopped-In-Ranking-Lawsuit-Against-Google (“So what is it, objective
or subjective, or argue what’s most convenient . . . [?]”).
874 MINNESOTA LAW REVIEW [98:868
bumps] or [DIY subwoofer]—expresses a desire to become
21
better-informed on a specific subject. The queries, and the de-
sires, are as diverse as the range of human experience. The
search results that come back are a personally customized mix
of websites; if the process is working well, they are uniquely
relevant to the user’s unique interests. This is profoundly good
for individual autonomy, and the law ought to promote it. Good
search policy would put users first.
From the user’s perspective, a search engine is not primari-
ly a conduit or an editor. Instead, it is a trusted advisor. It lis-
tens to a user’s description of her goals in the form of a search
query, performs research on her behalf, uses its expert judg-
ment to sift through what it has learned, and reports back to
her with recommendations on which websites to visit and
which ones to ignore. This point of view harmonizes the conduit
and editor theories by incorporating insights from both. A
search engine connects websites to users and it exercises dis-
cretion in creating its results. The two functions are insepara-
ble because they are both aspects of advising search users
about websites.
There are threads of the advisor theory throughout the ex-
isting debates on search. But because these debates have tradi-
tionally been understood as a series of binary choices—conduit
or editor, objective or subjective, for Google or against it—their
significance has been discounted. Users’ interests are present,
but only rhetorically, as a justification for siding with websites
or with search engines. Commentators simply assume that the
question to be answered is whether treating search engines as
conduits or editors is better for users in the long run. That both
of these options might sell users short has not previously been
suggested.
If we are determined, as we should be, to put search users
first, law can do two things for them. It can promote access to
search by enabling users to draw on the aid of search engines,
and it can promote loyalty in search by preventing search en-
gines from misleading users. Access responds to the conduit
theory: the search engine owes nothing to websites struggling
to be heard; what matters is the user’s ability to select among
websites, which necessarily includes ignoring most of them
most of the time. And loyalty responds to the editor theory: a
search result is not a product the user consumes for its own sa-
21. Except where a citation is provided, all examples of search queries in
this Article are fictional.
2014] SPEECH ENGINES 875
ke; it is useful only as a way to find the websites whose speech
the user really values.
Access and loyalty provide fresh insights into numerous le-
gal and regulatory debates about search. Take a search bias
claim like Foundem’s: that Google unfairly lowered its search
22
ranking. On the conduit theory, which says that search re-
sults are objective, Foundem’s claim should succeed, as long as
Foundem is right that its website really is better than the al-
ternatives. On the editor theory, which says that search results
are subjective, Foundem’s claim is doomed at the outset: Google
is categorically free to express its own opinion of websites.
On the advisor theory, matters are more nuanced. From
users’ point of view, website quality is subjective; no two users
will have quite the same preferences. You say tomato.com; I say
tomato.org. In a search for [tomato], either result could be
right. If we care about access, then no website ever has a right
to insist on top placement; if it did, it would override the pref-
erences of users who are looking for something else. But if we
care about loyalty, then Google is not yet off the hook. Search
engines systematically measure user satisfaction with search
results; they develop algorithms intended to return the results
23
users want. If Google shows tomato.com but knows that most
users would have preferred to see tomato.org, it disserves users
because it deceives them. The search results are not wrong in
an absolute sense, but they are dishonest in the context of
Google’s relationship to its users.
In fact, this is very close to the approach the FTC took
when it dismissed the search-bias portion of its investigation.
Google’s favoritism towards its own maps, flights, and other
vertical content was acceptable, the Commission wrote, because
“Google’s primary goal in introducing this content was to quick-
24
ly answer, and better satisfy, its users’ search queries . . . .”
On the conduit and editor theories, Google’s motives should
have been irrelevant: both theories focus on conduct, one to
condemn, the other to justify. But on the advisor theory, motive
is crucial, because it is the intent to harm users that makes the
ranking disloyal and thus actionable. The FTC did not explain
why its analysis properly turned on Google’s motives; the advi-
sor theory supplies the missing explanation.
22. This argument is presented in more detail infra Part IV.
23. See generally Steven Levy, How Google’s Algorithm Rules the Web,
WIRED, Mar. 2010, at 96.
24. Statement of the Fed. Trade Comm’n, supra note 5.
876 MINNESOTA LAW REVIEW [98:868
This Article presents, defends, and applies the advisor the-
ory of search. Part I is background. It provides a quick tech-
nical overview of how search engines work and a glossary of
important search terminology. Part II describes the conduit
and editor theories, showing how they embrace two fundamen-
tally incompatible visions of what search is and how to regulate
it. Part III resolves the tension between them by introducing
and defending the advisor theory. Part IV applies the advisor
theory to the search bias issue at the heart of the FTC’s inves-
tigation, concluding that the FTC was probably right to drop
the investigation without action. And Part V shows that the
advisor theory is useful more broadly, presenting four short
case studies of how it offers useful advice on other recurring
problems in search law.
I. HOW SEARCH ENGINES WORK
This Part briefly explains how modern search engines work
and how they display results to users. Readers who are familiar
with terms like “organic results” and “universal search” should
feel free to skip ahead to Part II.
If you type [learn spanish] into Google you will see
something like the following:
2014] SPEECH ENGINES 877
In the argot of search, this is a results page created based
25
on the query [learn spanish]. This particular query con-
26
tains two keywords or search terms. The displayed portion of
the results page shows three organic links and two sets of spon-
27
sored links (or search ads). Each of the results consists of a
link back to the website (underlined in blue), a text version of
the website’s online address (in green), and two lines of text ex-
cerpted from the website or supplied by it (in black). For image
searches, the excerpt is a thumbnail of the image; for books it is
an eighth-of-a-page snippet; a video is usually represented by a
single frame. The ordinal position of a result is called its rank-
28
ing.
The organic results are generated through a three-step
29
process. First, Google’s computers crawl webpages and other
sources to learn about what information they contain and how
30
they link to each other. Second, Google uses complex and
time-consuming algorithms to analyze the pages and their rela-
31
tionships. These estimates are based on hundreds of signals
that assess pages’ importance and their relevance to particular
32
search terms. Examples of signals include whether a website
is commercial or educational, how recently a webpage was up-
dated, and whether a letter followed by a period might be a
33
middle initial. The third step comes in response to the user’s
query: Google consults its tables of signals, identifies those
webpages that score highly for the query, and then displays
34
them in descending order of relevance. All of this is complete-
25. See Levy, supra note 23 (explaining queries). See generally JOHN
BATTELLE, THE SEARCH: HOW GOOGLE AND ITS RIVALS REWROTE THE RULES
OF BUSINESS AND TRANSFORMED OUR CULTURE (2005) (giving history of web
search); STEVEN LEVY, IN THE PLEX: HOW GOOGLE THINKS, WORKS, AND
SHAPES OUR LIVES (2011) (bringing history up to date through 2011).
26. See Levy, supra note 23.
27. See Grimmelmann, supra note 1, at 23.
28. See Pasquale, supra note 11. Confusingly, a “top” ranking means visu-
ally near the top of the first page of results, and a “low” ranking is anything
else—but it is actually the “low” rankings that have numerically greater ordi-
nal rank when the results are ordered “first,” “second,” and so on.
29. See BATTELLE, supra note 25, at 20.
30. Id.
31. Id. at 22.
32. Levy, supra note 23.
33. Id.
34. Id.
878 MINNESOTA LAW REVIEW [98:868
ly standard for modern search engines; only the specific signals
35
differentiate one search engine from another.
36
Traditionally, there were two types of search engines.
“General” search engines indexed the entire web; “vertical”
search engines narrowed their focus to a specific type of con-
37 38
tent, such as movies, hotel bookings, African-American
39 40
themes, or product comparisons. Google initially expanded
into vertical search with specialized local search, news search,
41
and image search, each with its own URL. In May 2007, the
company took an important step to break down these distinc-
tions. It launched Universal Search, which “blend[ed] content
from Images, Maps, Books, Video, and News into [Google’s] web
42
results.” Here is an example showing a restaurant-themed
search for [hamburgers in topeka ks]. The top three re-
sults are standard general web results from the third-party
websites Urbanspoon and Topeka.net. But they are followed by
local search results from Google’s vertical local-search engine,
and to the right of the main column is a map from Google Maps
showing the locations of those restaurants.
35. Id.
36. See Statement of the Fed. Trade Comm’n, supra note 5 (“General pur-
pose search engines are distinct from ‘vertical’ search engines, which focus on
narrowly defined categories of content such as shopping or travel.”).
37. See, e.g., Advanced Search, IMDB, http://www.imdb.com/search/ (last
visited Nov. 27, 2013).
38. See, e.g., HOTELS.COM, http://www.hotels.com (last visited Nov. 27,
2013).
39. See, e.g., AFROROOTS, http://www.afroroots.com (last visited Nov. 27,
2013).
40. See, e.g., PRICEGRABBER, http://www.pricegrabber.com (last visited
Nov. 27, 2013).
41. See, e.g., News, GOOGLE, http://news.google.com (last visited Nov. 27,
2013).
42. See Marissa Mayer, Universal Search: The Best Answer Is Still the
Best Answer, GOOGLE OFFICIAL BLOG (May 16, 2007), http://googleblog
.blogspot.com/2007/05/universal-search-best-answer-is-still.html (publicizing
release of Google’s universal search).
2014] SPEECH ENGINES 879
Google is hardly alone in starting down this road. Mi-
crosoft’s Bing, for example, has many of the same categories of
vertical results as Google does. But Google’s powerful posi-
43
tion—it has 67% of the United States search market and up-
44
wards of 90% in some European countries —has given its move
to universal search a special urgency.
II. THE CONDUIT AND EDITOR THEORIES
Everyone claims to have users’ interests at heart, and yet
the pro- and anti-Google camps are at loggerheads over how
best to help them. The explanation is the powerful gravitation-
al pull of the conduit and editor theories of search. Commenta-
tors who start off talking about what would be best for users
find themselves drawn—often without even realizing it—to one
of these decidedly non-user-centric theories of search. To un-
derstand what a truly user-centric theory of search would look
like, therefore, it is necessary to start by bringing out the im-
plicit assumptions made by these other theories.
The plan of this Part is simple: Section A presents the con-
duit theory and Section B the editor theory. Section C then
43. See ComScore Releases January 2013 U.S. Search Engine Rankings,
COMSCORE (Feb. 13, 2013), http://www.comscore.com/Insights/Press_Releases/
2013/2/comScore_Releases_January_2013_U.S._Search_Engine_Rankings.
44. Paul Geitner, Google Moves Toward Settlement of European Antitrust
Investigation, N.Y. TIMES, July 25, 2012, at B3.
880 MINNESOTA LAW REVIEW [98:868
highlights three important and interesting contrasts between
the two.
A. THE CONDUIT THEORY
When you write about search, you get the most interesting
emails. For example:
On a contingency fees Basis, I want to sue Google, who, using its
dominant position (and through Googlebot actions,
regularly registered on my Web Sitemeter) censors, constantly, de-
liberately and vigorously, Texts and Images of my (un-harmful) Web-
45
site . . . as it can be easily proved.
This brief plea for help is the conduit theory in a nutshell: the
law should prevent Google from using its “dominant position”
to “censor” websites.
Websites and their Google nightmares are at the heart of
the conduit theory. Sometimes, the harms are economic: the
French legal search engine eJustice lost customers and adver-
tising revenue after being demoted in Google’s rankings; it sued
46
for €295 million. But there are just as many stories in which
speech is at stake; Christopher Langdon sued Google for refus-
ing to let him advertise his websites NCJusticeFraud.com and
47
ChinaIsEvil.com.
The scholars who sympathize with these websites draw on
the well-established tradition in free-speech theory that speak-
ers should have an affirmative right of access to the mass me-
48
dia. They argue that speakers can effectively reach the public
only with the media’s help; where that help is withheld, the re-
49
sult is private censorship. Telecommunications law’s long his-
50
tory of nondiscrimination rules, from the treatment of tele-
45. Email to James Grimmelmann (Jan. 26, 2013) (on file with author).
46. See INITIATIVE FOR A COMPETITIVE ONLINE MARKETPLACE, GOOGLE
UNDER THE ANTITRUST MICROSCOPE 12–13 (2011), available at http://ec
.europa.eu/information_society/newsroom/cf/pillar.cfm.
47. See Langdon v. Google Inc., 474 F. Supp. 2d. 622, 626 (D. Del. 2007).
48. The urtext for this tradition is Jerome A. Barron, Access to the Press—
A New First Amendment Right, 80 HARV. L. REV. 1641 (1967). For modern
commentary on the tradition, see generally Marjorie Heins & Eric M. Freed-
man, Reclaiming the First Amendment: Constitutional Theories of Media Re-
form, 35 HOFSTRA L. REV. 917 (2007) (symposium issue on the 40th anniver-
sary of Barron’s original article); Symposium, Access to the Media: 1967 to
2007 and Beyond: A Symposium Honoring Jerome A. Barron’s Path-Breaking
Article, 76 GEO. WASH. L. REV. 819 (2008) (same).
49. See Jack M. Balkin, Media Access: A Question of Design, 76 GEO.
WASH. L. REV. 933, 935 (2008).
50. See generally Daniel A. Lyons, Net Neutrality and Nondiscrimination
Norms in Telecommunications, 54 ARIZ. L. REV. 1029 (2012); Tim Wu, Why
2014] SPEECH ENGINES 881
51
phone and telegraph as common carriers to the recent push
52
for network neutrality, embody this philosophy. So did right-
53
of-reply statutes and the Federal Communications Commis-
54
sion’s late fairness doctrine, both of which compelled mass
55
media to present opposing viewpoints. In each case, the medi-
um is regarded as a conduit: it exists to carry the speech of oth-
56
ers.
With the rise of the Internet and of Google, scholars have
57
extended this argument to search engines. The argument re-
quires one significant modification, because search engines are
58
not “means of speech” like printing presses or cable networks.
59
Instead, they are “selection intermediaries” that direct users
60
to one information provider or another. Because of their role,
they have immense power to choose which speakers are found
61
and which speakers are sent “to the back row of the arena.”
Have a Telecommunications Law? Anti-Discrimination Norms in Communica-
tions, 5 J. TELECOMM. & HIGH TECH. L. 15 (2006).
51. See 47 U.S.C. § 202 (2006) (prohibiting “any unjust or unreasonable
discrimination” by common carriers).
52. See FED. COMMC’NS COMM’N, REPORT AND ORDER: PRESERVING THE
FREE AND OPEN INTERNET, No. 10-201 (2010).
53. See Miami Herald Publ’g Co. v. Tornillo, 418 U.S. 241, 244 (1974) (de-
scribing Florida’s right-of-reply statute).
54. See Red Lion Broad. Co. v. FCC, 395 U.S. 367, 373–86 (1969) (describ-
ing history of fairness doctrine).
55. See generally Television Assignments, 41 F.C.C. 148, 158–67 (1952)
(reserving television channels for educational broadcasting); Turner Broad.
Sys., Inc. v. FCC, 520 U.S. 180 (1997) (upholding statute requiring cable oper-
ators to carry certain local broadcast television stations).
56. See Oren Bracha & Frank Pasquale, Federal Search Commission?: Ac-
cess, Fairness, and Accountability in the Law of Search, 93 CORNELL L. REV.
1149, 1199 (2008) (“[S]earch engine rankings play a central instrumental role
in facilitating effective speech by others.”); Emily B. Laidlaw, Private Power,
Public Interest: An Examination of Search Engine Accountability, 17 J.L. INFO
& TECH. 113, 122 (2008) (“The media’s core role is to publish.”); see also Note,
Against Search Engine Volition, 18 ALB. L.J. SCI. & TECH. 205, 223 (2008)
(“[T]he search engine’s effective function is to serve as a conduit between third
parties.”).
57. See, e.g., Lucas D. Introna & Helen Nissenbaum, Shaping the Web:
Why the Politics of Search Engines Matter, 16 INFO. SOC’Y 169, 169–70 (2000)
(drawing parallels between debates about other media and search engines on
the Internet).
58. Chandler, supra note 10, at 1097.
59. Id.
60. See Laidlaw, supra note 56, at 123–26.
61. Jonathan Rosenberg, From the Height of This Place, GOOGLE OFFICIAL
BLOG (Feb. 16, 2009), http://googleblog.blogspot.com/2009/02/from-height-of
-this-place.html. For further examples of this discourse, see, e.g., Niva Elkin-
Koren, Let the Crawlers Crawl: On Virtual Gatekeepers and the Right to Ex-
882 MINNESOTA LAW REVIEW [98:868
Since search engines have the same practical power as tradi-
tional mass media to shape public discourse, goes the argu-
ment, they should be subject to the same scrutiny and perhaps
62
to the same regulations. Some scholars have argued that web-
sites should have an affirmative right to be included in search
63
engine indices. Most recently, some observers have proposed
64
“search neutrality” rules by analogy to network neutrality.
They all agree that the legal system should ensure that a di-
verse array of information providers can be found through
65
search engines —that the search engine is a conduit for their
66
ideas.
clude Indexing, 26 U. DAYTON L. REV. 179, 183 (2001) (“[F]ocal point of control
. . . .”); Editorial, The Google Algorithm, N.Y. TIMES, July 15, 2010, at A30
(“[T]hey can break the business of a Web site that is pushed down the rank-
ings.”).
62. See, e.g., Elkin-Koren, supra note 61, at 184 (comparing Internet users
to television viewers); Laidlaw, supra note 56, at 123–37 (“Within the regula-
tory models in traditional media, the Internet adopts some of all three func-
tions.”). But see, e.g., Bracha & Pasquale, supra note 56, at 1157 (“Internet
communication does not possess the characteristics that supported broad regu-
lation of broadcast media.”).
63. Pasquale, supra note 11, at 135–36 (2006) (proposing right of reply to
harmful search results with asterisk); see, e.g., DAWN NUNZIATO, VIRTUAL
FREEDOM 14–17, 149–51 (2009); Chandler, supra note 10, at 1117–25; Dawn
Nunziato, Death of the Public Forum in Cyberspace, 20 BERK. TECH. L.J. 1115,
1123–25 (2005). But see James Grimmelmann, Don’t Censor Search, 117 YALE
L.J. POCKET PART 48 (2007), http://Yalelawjournal.org/images/pdfs/
582.pdf (critiquing asterisk); Frank Pasquale, Asterisk Revisited: Debating a
Right of Reply on Search Results, 3 J. BUS & TECH. L. 61 (2008) (defending as-
terisk).
64. Nate Anderson, Search Neutrality? How Google Became a Neutrality
Target, ARS TECHNICA (Apr 29, 2010, 9:32 PM), http://arstechnica.com/tech
-policy/2010/04/search-neutrality-google-becomes-neutraliy/ (describing history
of idea); see also John Blevins, The New Scarcity: A First Amendment Frame-
work for Regulating Access to Digital Media Platforms, 79 TENN. L. REV. 353
(2012); Jeremy Jarosch, Novel “Neutrality” Claims Against Internet Platforms:
A Reasonable Framework for Initial Scrutiny, 59 CLEV. ST. L. REV. 537 (2011);
Andrew Odlyzko, Network Neutrality, Search Neutrality, and the Never-
Ending Conflict Between Efficiency and Fairness in Markets, 8 REV. NETWORK
ECON. 40 (2009); Raff, supra note 2. The Foundem-founded website
SearchNeutrality.org offers a useful roundup of links and commentary on the
pro-search-neutrality side. SEARCHNEUTRALITY.ORG: A FOUNDEM INITIATIVE,
http://www.searchneutrality.org (last visited Nov. 27, 2013). For a contrasting
point of view, see James Grimmelmann, Some Skepticism About Search Neu-
trality, in THE NEXT DIGITAL DECADE: ESSAYS ON THE FUTURE OF THE INTER-
NET 435–47 (Berin Szoka & Adam Marcus eds., 2010).
65. Most ambitiously, some commentators proposed randomization, in
which sites would be randomly promoted in search engine rankings. See, e.g.,
Sandeep Pandey et al., Shuffling a Stacked Deck: The Case for Partially Ran-
domized Ranking of Search Results, PROC. 31ST CONF. ON VERY LARGE DATA-
2014] SPEECH ENGINES 883
Three recurring metaphors illustrate how the conduit theo-
ry thinks about search: “‘maps,’” ‘“traffic,’” and ‘“bias.’” Scholars
67
who say that search engines “create a map of the Web” or that
68
Google is “the main map to the information highway” appeal
69
to an ideal of accuracy and objectivity. Oren Bracha and
Frank Pasquale propose that search results be treated like
70
maps, and the intuitive justification is simple. A good map
shouldn’t say that there’s a bridge where there isn’t one in real
71
life.
The related traffic metaphor is even sharper. Saying that a
72
search engine delivers “traffic” to websites implies that search
73
is a kind of transportation infrastructure. When Google “di-
vert[s] traffic,” the metaphor suggests an unwanted detour, like
74
orange cones forcing users off the highway at the Google exit.
BASES (2005), available at http://www.cs.cmu.edu/~spandey/publications/
randomRanking-vldb.pdf.
66. See, e.g., Bracha & Pasquale, supra note 56, at 1192 (“[S]earch engines
. . . portray[] themselves as passive conduits.”); Laidlaw, supra note 56, at 124
(“[S]earch engines are now the portals through which the information on the
Internet is experienced.”).
67. Introna & Nissenbaum, supra note 57, at 171.
68. The Google Algorithm, supra note 61.
69. See Grimmelmann, supra note 1, at 31–33.
70. Bracha & Pasquale, supra note 56, at 1194 n.238. The authors also
compare search results to directory listings, concluding, “It is hard to conceive
of a phone book as embodying any constitutionally protected message.” Id.
71. Of course, maps can be inaccurate and intensely political; there are
sometimes sharp controversies over what maps do and do not show. See gener-
ally MARK MONMONIER, HOW TO LIE WITH MAPS (2d ed. 1996). Google has had
a few unfortunate encounters with this phenomenon. See Frank Jacobs, The
First Google Maps War, N.Y. TIMES OPINIONATOR (Feb. 28, 2012, 12:30 PM),
http://opinionator.blogs.nytimes.com/2012/02/28/the-first-google-maps-war/
(discussing tension between Nicaragua and Costa Rica over incorrect Google
Maps depiction of disputed frontier territory); see also Rosenberg v. Harwood,
No. 100916536, 2011 BL 333199, at *4–5 (Utah Dist. Ct. May 27, 2011) (hold-
ing that Google owed no duty to warn Google Maps user that highway lacked
sidewalks); infra Part IV.B (discussing tort liability for publishers of inaccu-
rate maps).
72. See, e.g., CONSUMER WATCHDOG, TRAFFIC REPORT: HOW GOOGLE IS
SQUEEZING OUT COMPETITORS AND MUSCLING INTO NEW MARKETS 15 (2010);
INITIATIVES FOR A COMPETITIVE MARKET, supra note 46, at 8.
73. Cf. BRETT M. FRISCHMANN, INFRASTRUCTURE: THE VALUE OF SHARED
SOCIAL RESOURCES 359–60 (2012) (discussing Google Books corpus and set-
tlement as possible infrastructural resources).
74. See Alex Barker, Antitrust Chief Holds All the Aces, FIN. TIMES (Lon-
don), Jan. 11, 2013, at 17 (quoting Joaquín Almunia, European Commissioner
for Competition); see also, e.g., ADAM RAFF & SHIVAUN RAFF, PENALTIES,
SELF-PREFERENCING, AND PANDA 1 (2011) (“Google can exploit its gatekeeper
status to commandeer a substantial proportion of the traffic of almost any
884 MINNESOTA LAW REVIEW [98:868
It also downplays any speech element in search; driving is a
75
form of conduct, not communication. As Bracha and Pasquale
put it, “[r]ankings are functional rather than dialogical expres-
76
sions.”
The bias metaphor describes what happens when search
77
engines fall short of this ideal. Friedman and Nissenbaum de-
fine “biased” computer systems as ones that “systematically and
unfairly discriminate against certain individuals or groups of
78
individuals in favor of others.” As applied to search engines,
the idea is that an engine may skew its results in a way that
79
imposes its own viewpoint on users. These biases could be
large and subtle—towards commercial content or popular
80
sites —or they could be specific and targeted to advance the
website or industry sector it chooses.”); Grimmelmann, supra note 1, at 27–30.
75. Cf. Bracha & Pasquale, supra note 56, at 1197 (comparing search re-
sult to seller offering “three alternative products in response to a buyer’s in-
quiry”).
76. Id. at 1198.
77. The metaphor is pervasive in academic critiques of Google. For exam-
ples of papers that use the metaphor in the title, see Benjamin Edelman, Bias
in Search Results? Diagnosis and Response, 7 INDIAN J.L. & TECH. 16 (2011);
Goldman, supra note 12; Joshua G. Hazan, Stop Being Evil: A Proposal for
Unbiased Google Search, 111 MICH. L. REV. 789 (2013); Introna &
Nissenbaum, supra note 57; Alejandro Diaz, Through the Google Goggles: So-
ciopolitical Bias in Search Engine Design (May 2005) (unpublished under-
graduate thesis, Stanford University), available at http://epl.scu.edu/~
stsvalues/readings/diaz_thesis_final.pdf.
The most remarkable—and embarrassing for Google—use of the term
must be by Google’s founders. See SERGEY BRIN & LAWRENCE PAGE, THE
ANATOMY OF A LARGE-SCALE HYPTERTEXTUAL WEB SEARCH ENGINE app. A
(1998), available at http://ilpubs.stanford.edu:8090/361/1/1998-8.pdf (“Since it
is very difficult even for experts to evaluate search engines, search engine bias
is particularly insidious.”). Related metaphors get at the idea of unfairness by
treating search as a game. See The Power of Google: Serving Consumers or
Threatening Competition?: Hearing Before the Subcomm. on Antitrust, Compe-
tition Policy, & Consumer Rights of the S. Comm. On the Judiciary, 112th
Cong. 28 (2011) (testimony of Jeremy Stoppelman, CEO, Yelp! Inc.) [hereinaf-
ter Stoppelman Testimony]; Jeffrey Katz, Op-Ed., Google’s Monopoly and In-
ternet Freedom, WALL ST. J., June 8, 2012, at A15, available at http://online
.wsj.com/news/articles/SB10001424052702303830204577448792246251470
(stating that Google has “stacked the deck”).
78. Batya Friedman & Helen Nissenbaum, Bias in Computer Systems, 14
ACM TRANSACTIONS ON INFO. SYS. 330, 332 (1996).
79. See, Chandler, supra note 10, at 1103; James Grimmelmann, Infor-
mation Policy for the Library of Babel, 3 J. BUS. & TECH. L. 29, 39 (2008)
(“Search engines can play favorites.”); Introna & Nissenbaum, supra note 57,
at 176 (“skewed”).
80. Cf. Urs Gasser, Regulating Search Engines: Taking Stock and Looking
Ahead, 8 YALE J.L. & TECH 201, 228–29 (2006) (discussing “diversity” of in-
formation sources as a policy goal for search). See generally MATTHEW
2014] SPEECH ENGINES 885
81
search engine’s own commercial interests. But whatever form
82
it takes, bias is bad.
B. THE EDITOR THEORY
Some of Google’s defenders have a surprising response to
search bias: they embrace it. Blogger Mike Masnick writes,
“[T]here’s no such thing as ‘neutrality’ in search, because any
83
ranking is biased by what the search engine thinks is best.” Or,
as Eric Goldman argues:
Search engines are media companies. Like other media companies,
search engines make editorial choices designed to satisfy their audi-
ence. These choices systematically favor certain types of content over
others, producing a phenomenon called “search engine bias.”
Search engine bias sounds scary, but . . . such bias is both neces-
84
sary and desirable.
And that’s the editor theory in a nutshell: search engines are
“media companies” that make “editorial choices” about what to
publish.
In the words of a Google engineer, “In some sense when
people come to Google, that’s exactly what they’re asking for—
85
our editorial judgment.” Editor theorists agree that search re-
sults are “editorial judgments” about which websites might be
86
of interest to users. Search engines are editors that pick and
HINDMAN, THE MYTH OF DIGITAL DEMOCRACY 54–57 (2009) (discussing
“Googelearchy” in which popular sites become even more popular); Laidlaw,
supra note 56, at 129; Diaz, supra note 77, at 62–94.
81. See generally Bracha & Pasquale, supra note 56, at 1170; Edelman,
supra note 77, at 19 (describing informal experiment to show pro-network
neutrality bias in search results of network neutrality supporter Google);
Friedman & Nissenbaum, supra note 78, at 330–31 (describing biases in com-
puterized reservation systems favoring the systems’ airline owners).
82. See BEUC, FAIR INTERNET SEARCH: REMEDIES IN GOOGLE CASE 5
(2013), available at http://www.beuc.org/custom/2013-00211-01-E.pdf (“Google
must use an objective, non-discriminatory mechanism to rank and display all
search results, including any links to Google products.”).
83. Mike Masnick, A Recommendation Is Not the Same as Corruption,
TECHDIRT (June 21, 2010, 8:07 PM), http://www.techdirt.com/articles/
20100621/0355239887.shtml.
84. Goldman, supra note 12, at 189; see also Christopher S. Yoo, Free
Speech and the Myth of the Internet as an Unintermediated Experience, 78
GEO. WASH. L. REV. 697, 708 (2010) (“It is thus hard to see how to make sense
of criticisms that search engine results are ‘biased’ when bias is the very es-
sence of the enterprise.”).
85. Steven Levy, TED 2011: The ‘Panda’ That Hates Farms: A Q&A with
Google’s Top Search Engineers, WIRED NEWS (Mar. 3, 2011, 1:36 PM), http://
www.wired.com/business/2011/03/the-panda-that-hates-farms/all/.
86. See Eric Goldman, Revisiting Search Engine Bias, 38 WM. MITCHELL
886 MINNESOTA LAW REVIEW [98:868
choose among preexisting materials to generate a new presen-
87
tation. Making the necessary choices requires the exercise of
88
discretion and judgment. Google even holds a patent on a
“[s]ystem and method for supporting editorial opinion in the
89
ranking of search results.”
The editor theory has its own long and distinguished tradi-
90
tion in free speech law and theory. The press is so central to
the First Amendment that it is called out by name; the United
States Reports are stuffed with encomia to the democratic con-
91
tributions of editors and publishers. Editors must be free to
select and present unpopular and controversial viewpoints; the
government is forbidden to interfere with their exercise of pro-
fessional judgment.
Exhibit A for the editor theory is an analogy between
92
Google and newspapers, most often the New York Times. In
L. REV. 96, 106 (2011) (“editorial discretion”); Goldman, supra note 12, at 189
(“editorial choices”); id. at 190 (“editorial judgments”); Volokh & Falk, supra
note 13, passim; Brown & Davidson, supra note 6 (“editorial judgment”); Dan-
ny Sullivan, The New York Times Algorithm & Why It Needs Government Reg-
ulation, SEARCH ENGINE LAND (Jul. 15, 2010, 2:07 PM), http://
searchengineland.com/regulating-the-new-york-times-46521 (“editorial judg-
ment”).
87. See, e.g., Volokh & Falk, supra note 13, at 890 (“[Search results] are
collections of facts that are organized and sorted using the judgment embodied
in the engines’ algorithms . . . .”); cf. 17 U.S.C. § 101 (2012) (“A ‘compilation’ is
a [copyrightable] work formed by the collection and assembling of preexisting
materials or of data that are selected, coordinated, or arranged in such a way
that the resulting work as a whole constitutes an original work of author-
ship.”).
88. See Volokh & Falk, supra note 13, at 891 (“The First Amendment pro-
tects the decisions to include or exclude others’ content, based on the speakers’
exercise of their judgment . . . .”).
89. U.S. Patent No. 7,096,214 (filed Dec. 13, 2000).
90. See, e.g., Randall Bezanson, The Developing Law of Editorial Judg-
ment, 78 NEB. L. REV. 754, 756 (1999) (reviewing role of press in free speech
doctrine and theory, with particular focus on “the press’s central instrument,
editorial judgment, and its main claim to constitutional protection, editorial
freedom”).
91. See Pell v. Procunier, 417 U.S. 817, 832 (1972) (“The constitutional
guarantee of a free press assures the maintenance of our political system and
an open society and secures the paramount public interest in a free flow of in-
formation to the people concerning public officials.” (internal quotation marks
omitted)); Associated Press v. United States, 326 U.S. 1, 20 (1944) (“That
Amendment rests on the assumption that the widest possible dissemination of
information from diverse and antagonistic sources is essential to the welfare of
the public, that a free press is a condition of a free society.”).
92. See Schmidt Testimony, supra note 16 (“Just as a government panel
could not dictate to the New York Times, the Drudge Report, or the Huffington
Post what stories they could publish on their websites without infringing their
2014] SPEECH ENGINES 887
response to a Times editorial calling for greater scrutiny of
93
Google’s search results, search industry analyst Danny Sulli-
van wrote a scathing response turning the editorial’s argu-
94
ments back on the newspaper. “When the New York Times ed-
itorial staff tweaks its supersecret algorithm behind what to
cover and exactly how to cover a story—as it does hundreds of
times a day—it can break a business that is pushed down in
95
coverage or not covered at all.” The argument for regulating
the Times’ editorial meetings is meant to be absurd, and there-
by to illustrate the absurdity of the Times’ argument for regu-
96
lating search rankings.
A second characteristic trope of the editor theory is that
search is hard. Commentators and advocates describe the diffi-
culty of the search ranking process: the unfathomable number
97
of webpages Google indexes, the number of distinct signals on
98 99
which it relies, the number of changes it makes a year, the
extensive work that goes into assessing and improving the re-
100
sults. These points are directed to showing that regulation of
freedom of speech, so too would government-mandated results likely violate
Google’s freedom of speech.”); Goldman, supra note 12, at 189 (“Search engines
are media companies.”); Volokh & Falk, supra note 13, at 884–85 (comparing
search engines to newspapers); id. at 888 (“[S]earch engine companies are
rightly seen as media enterprises, much as the New York Times Company or
CNN are media enterprises.”); Brown & Davidson, supra note 6 (“But search
engines need to make choices about what results are most relevant to a query,
just as a news editor must decide which stories deserve to be on the front
page.”); Sullivan, supra note 86; cf. Chandler, supra note 10, at 1126–29 (ac-
knowledging similarity between selection intermediaries and newspapers but
arguing for constitutionality of transparency and anti-blocking rules).
93. The Google Algorithm, supra note 61.
94. Sullivan, supra note 86.
95. Id.
96. Id. (“Suffice to say, the editorial staff of the New York Times would
scream bloody murder if anyone suggested government oversight of its own
editorial processes.”).
97. See James Grimmelmann, The Google Dilemma, 53 N.Y.L. SCH. L.
REV. 939, 940 (2009); Volokh & Falk, supra note 13, at 891 (“Search engines
are vastly more selective . . . .”).
98. See, e.g., Schmidt Testimony, supra note 16. See generally Levy, supra
note 23 (describing signals).
99. Schmidt Testimony, supra note 16. For examples, see Pandu Nayak,
Search Quality Highlights: 65 Changes for August and September, GOOGLE
INSIDE SEARCH (Oct. 4, 2012, 9:12 AM), http://insidesearch.blogspot.com/2012/
10/search-quality-highlights-65-changes.html (listing 65 algorithmic changes
in 61 days).
100. See, e.g., Marissa Mayer, Do Not Neutralize the Web’s Endless Search,
FIN. TIMES (London) (July 15, 2010, 11:19 PM), http://www.ft.com/intl/cms/s/
0/0458b1a4-8f78-11df-8df0-00144feab49a.html#axzz2hj5r3qgT (“Yet searching
888 MINNESOTA LAW REVIEW [98:868
101
search results would be futile, but they are also intended to
102
demonstrate the human judgment involved. In March 2012,
Google released a video of a short segment of its weekly search
quality meeting, showing its engineers spending eight minutes
debating, with extensive empirical data, how to choose which
103
words to spell-check in long search queries. The resemblance
to a newspaper editorial meeting cannot have been lost on
Google’s public-relations team.
A final trope of the editor theory is innovation in search
technology. Search has progressed far beyond the “ten blue
links” of a decade ago to a paradigm of universal search incor-
104
porating structured vertical results. One point of emphasiz-
ing this evolution is to argue that Google’s changes reflect in-
dustry-wide advances in how results are organized and
105
presented to users, not nefarious motives unique to Google.
Another reason is to demonstrate the existence of vigorous
competition in the industry, so that Google can defend its posi-
106
tion only through vigorous innovation in improving search.
And a third is to argue that any regulation of search results
107
would inhibit future advances. All of these arguments take a
romantic view of the search engineer: he (or she, but usually
the web has never been more complex.”).
101. See, e.g., Marvin Ammori & Luke Pelican, Competitors’ Proposed Rem-
edies for Search Bias: Search “Neutrality” and Other Proposals, 15 J. INTER-
NET L., May 2012, at 1, 14–15.
102. See, e.g., Goldman, supra note 12, at 190; Volokh & Falk, supra note
13, at 888 (“[T]he computer algorithms that produce search engine output are
written by humans.”).
103. Google, Search Quality Meeting: Spelling for Long Queries (Annotat-
ed), YOUTUBE (Mar. 12, 2012), http://www.youtube.com/watch?&v=
JtRJXnXgE-A; see also Schmidt Testimony, supra note 16, at 3.
104. See Goldman, Revisiting Search Engine Bias, supra note 86, at 102–
05. “Ten blue links” is an informal search-engine industry term that refers to
the sparse look of a circa-2000 search engine results page: ten results, each
with a hyperlink in blue, and some descriptive text under each. See generally
Romy Kohavi, Ten Blue Links No More: Dynamic Page Sizing, BING BLOGS
(Apr. 24, 2013), http://www.bing.com/blogs/site_blogs/b/searchquality/archive/
2013/04/24/ten-blue-links-no-more-dynamic-page-sizing.aspx (describing the
ten blue links and the decreasing utility of such a format). The term’s empha-
sis is on what it excludes: images, maps, oneboxes, and other ways of present-
ing information. Id.
105. See, e.g., Ammori & Pelican, supra note 101, at 10–11.
106. See Schmidt Testimony, supra note 16, at 3–4; cf. Sullivan, supra note
86 (describing Yahoo!’s fall from dominance of the web).
107. See, e.g., Ammori & Pelican, supra note 101, at 19–20; Goldman, su-
pra note 12, at 197–98; Grimmelmann, supra note 63, at 50; Mayer, supra
note 100.
2014] SPEECH ENGINES 889
he) is a creative technical genius whose talents society should
108
harness by respecting his freedom to innovate.
C. THREE POINTS OF DISAGREEMENT
The most fundamental difference between the conduit and
editor theories is the way they think about speech. The conduit
theory focuses on what search does; the editor theory on what
search says. On the conduit theory, a search engine is a medi-
um, and as a medium it has little or no speech interest of its
own. It exists to help speakers reach audiences and the gov-
ernment should regulate it to that end. Conduit theorists argue
that the “expressive element [in search rankings] is overwhelm-
109
ingly minor and incidental.” Regulation to ensure that a
search engine provides access to speech by websites would not
110
interfere with any valuable speech by the search engine.
In sharp contrast, editor theorists focus not on websites’
speech, but on the search engine’s own speech. Take the news-
paper analogy. Newspapers show that editing is speaking, so if
the Times and other newspapers are fully protected by the First
111
Amendment, Google is too. The point is that the search en-
gine is cogitating and communicating in ways that entitle it,
normatively and legally, to the protections of free speech.
112
Search engines are speakers.
A second, related point of disagreement is that perennial
chestnut of search policy: whether search results are objective
or subjective. On the conduit theory, search results are, or
should be, objective. The assumption of the map metaphor is
that there is an underlying geography of information; an ideal
presentation would represent that geography with as little dis-
113
tortion as possible. A claim of bias implies the possibility of
108. Cf. Susan P. Crawford, Network Rules, 70 L. & CONTEMP. PROBS. 51,
53–54 (describing the rhetorical figure of the “romantic builder” who must be
free from governmental regulation to develop advanced communications net-
works).
109. Bracha & Pasquale, supra note 56, at 1193.
110. Id. at 1192; Chandler, supra note 10, at 1129.
111. Volokh & Falk, supra note 13, at 884; see also Langdon v. Google, 474
F. Supp. 2d 622, 629–30 (D. Del. 2007); cf. Hurley v. Irish-American Gay, Les-
bian & Bisexual Group, 515 U.S. 557, 570 (1995) (“[A]n edited compilation of
speech generated by other persons is a staple of most newspapers’ opinion
pages . . . .”).
112. See Volokh & Falk, supra note 13, at 884. But see Grimmelmann, su-
pra note 63, at 50 (“Search engines aren’t megaphones . . . .”).
113. See, e.g., Introna & Nissenbaum, supra note 57, at 172–73 (critiquing
assumption that particular technical processes used by search engines “are a
890 MINNESOTA LAW REVIEW [98:868
114
its absence. Search engine bias is deviation from an objective
115
ideal. Conduit theorists have turned Google’s own words
116
against it to argue that search is objective and impersonal.
They have particularly emphasized Google’s eagerness to dis-
claim legal responsibility for the information it links to and ex-
cerpts, presenting itself as a passive intermediary rather than
117
the source of that information.
On the editor theory, search results are inherently subjec-
tive because they express a search engine’s “opinion” about
118
websites. Where the conduit theory sees search rankings as
mechanical and objective, the editor theory describes them as
human and subjective, always uncertain and subject to de-
119
bate. Instead of decrying “bias,” the editor theory celebrates
it. Eric Goldman calls it “the unavoidable consequence of
120
search engines’ editorial control over their databases.” Others
go further, arguing that “bias” is a valuable expression of the
reliable indication of importance or relevance”).
114. Stoppelman Testimony, supra note 77, at 37 (“Google artificially pro-
motes its own properties regardless of merit.” (emphasis added)); Chandler,
supra note 10, at 1105 (“The key opportunity presented by the Internet is un-
filtered and essentially unbiased access to a vast quantity of speech”); Katz,
supra note 77 (“Google should provide consumers with access to the unbiased
search results it was once known for . . . .”).
115. See generally Edelman, supra note 77, at 21–23 (discussing construc-
tion of objective baseline for measuring bias). But see Joshua D. Wright, Defin-
ing and Measuring Search Bias: Some Preliminary Evidence (Geo. Mason L. &
Econ., Working Paper No. 12-14, 2011), available at http://papers.ssrn.com/
sol3/papers.cfm?abstract_id=2004649 (disputing Edelman’s methodology); cf.
Adam Raff & Shivaun Raff, Would the Real Search Neutrality Please Stand
Up, ORGZINE (May 30, 2012), http://zine.openrightsgroup.org/features/2012/
would-the-real-search-neutrality-please-stand-up (acknowledging that there is
no “right” answer to many searches but arguing that “any genuine pursuit of
the most relevant results must, by definition, preclude any form of arbitrary
discrimination.”).
116. Introna and Nissenbaum discuss and critique Yahoo!’s self-
presentations. Introna & Nissenbaum, supra note 57, at 172. For a general
discussion of the rhetoric used by search engineers, see Elizabeth Van
Couvering, Is Relevance Relevant? Market, Science, and War: Discourses of
Search Engine Quality, 12 J. COMPUTER-MEDIATED COMM. 866, 879 (2007).
117. See Bracha & Pasquale, supra note 56, at 1192; Richard Siklos, A
Struggle Over Dominance and Definition, N.Y. TIMES, Nov. 12, 2006, at B5
(quoting David Eun, a Google vice president, as saying, “I would say we’re a
conduit connecting our users with content and advertisers”).
118. Ammori & Pelican, supra note 101, at 17. For more extensive discus-
sion of the status of search results as opinions, see infra Part IV.
119. Volokh & Falk, supra note 13, at 884–85; Goldman, Search Engine Bi-
as, supra note 12, at 191 n.15; Ammori & Pelican, supra note 101 at 13, 19;
Mayer, supra note 100; Grimmelmann, supra note 64, at 443–44.
120. Goldman, supra note 12, at 195.
2014] SPEECH ENGINES 891
121
search engine’s own valuable opinions about content. For
122
them, neutrality is neither possible nor desirable.
A third point of disagreement is also revealing: competi-
tion. Conduit theorists describe the search market as concen-
123
trated and hard to break into, so that Google in particular
124
has substantial market power. Most Internet users find in-
125
formation through search engines. Search users are over-
whelmingly likely to follow links on the first page of results and
overwhelmingly more likely to follow links near the top of that
126
page. The result is that search engines are therefore “gate-
127
keepers” or “bottlenecks” on the Internet, so that websites are
128
utterly dependent on search engines. A website that drops in
121. Volokh & Falk, supra note 13, at 893.
122. Id. (“[S]ome hypothetical and undefined expectations of abstract objec-
tivity.”); Goldman, supra note 12, at 195 (“[S]earch engines simply cannot pas-
sively and neutrally redistribute third party content.”). Goldman, supra note
86, at 107 (“[N]eutral search engines . . . are entirely mythical.”).
123. INITIATIVE FOR A COMPETITIVE ONLINE MARKETPLACE, supra note 46,
at 30–31; Bracha & Pasquale, supra note 56, at 1180–86. See generally
FAIRSEARCH, FAIRSEARCH FACT SHEET: GOOGLE’S CONDUCT THREATENS CON-
SUMERS AND INNOVATION (2013), available at http://www.fairsearch.org/wp
-content/uploads/2011/06/Draft-Core-FairSearch-Fact-Sheet-051812.pdf (col-
lecting sources).
124. See, e.g., Mark R. Patterson, Google and Search Engine Market Power,
2013 HARV. J. L. & TECH. OCCASIONAL PAPER SERIES 1, 23 (2013).
125. See Laidlaw, supra note 56, at 126 (search engines are “indispensi-
ble”).
126. See Bernard J. Jansen et al., Determining the Informational, Naviga-
tional, and Transactional Intent of Web Queries, 44 INFO. PROCESSING &
MGMT. 1251 (2008); Bing Pan et al., In Google We Trust: Users’ Decisions on
Rank, Position, and Relevance, 12 J. COMPUTER MEDIATED COMM. 801, 806–07
(2007).
127. See, e.g., FAIRSEARCH, GOOGLE’S TRANSFORMATION FROM GATEWAY
TO GATEKEEPER: HOW GOOGLE’S EXCLUSIONARY AND ANTICOMPETITIVE CON-
DUCT RESTRICTS INNOVATION AND DECEIVES CONSUMERS 11 (2011); Bracha &
Pasquale, supra note 56, at 1150–51 (“Located at bottlenecks of the infor-
mation infrastructure, search engines exercise extraordinary control over data
flow in a largely decentralized network.”); Elkin-Koren, supra note 61, at 180
(“the new virtual gatekeepers of cyberspace”).
128. See, e.g., Grimmelmann, supra note 64, at 447 n.85 (collecting exam-
ples); INITIATIVE FOR A COMPETITIVE ONLINE MARKETPLACE, supra note 46, at
7–13 (collecting European examples); Introna & Nissenbaum, supra note 57 at
180; Elkin-Koren, supra note 61, at 184–85 (“If you are not listed in the search
results you are almost nonexistent on the web.”); Pasquale, Asterisk Revisited,
supra note 63, at 79 (“make-or-break power over internet-based businesses”);
The Power of Google: Serving Consumers or Threatening Competition?: Hear-
ing Before the Subcomm. on Antitrust of the S. Comm. on the Judiciary, 112th
Cong. 183 (2011) (written statement of Thomas O. Barnett, Partner, Coving-
ton & Burling) (“And to be found by consumers, particularly for new sites, a
website needs the ability to appear at or near the top of the results displayed
892 MINNESOTA LAW REVIEW [98:868
search rankings is “effectively ‘disappeared’ from the Inter-
129
net.”
The editor theory takes a very different view of competition
130
in the search market. In a phrase that Google has made a
131
mantra, “Competition is one click away.” It emphasizes the
132
existence of multiple search options, low user switching costs
133
to change search engines, consumers’ regular use of multiple
134 135
search engines, the entry of new specialized search engines,
and competition from other platforms like Twitter and Face-
136
book. These points are designed to emphasize that users have
broad and meaningful choice in how they find websites, and
that Google is far from the only way that websites can be
137
found. Indeed, it is common to see arguments that websites
should be careful not to become too dependent on the traffic
from any given search engine; if they do, they have only them-
138
selves to blame.
by a search engine.”).
129. See, e.g., Foundem’s Google Story, SEARCHNEUTRALITY.ORG (Aug. 18,
2009), http://www.searchneutrality.org/eu-launches-formal-investigation/
foundem-google-story; see also Introna & Nissenbaum, supra note 57, at 180
(noting that search engines wield power over websites in “making others, es-
sentially, disappear”). A related idea is that Google is a “killer.” See Katz, su-
pra note 77 (“brand killer”); Stoppelman Testimony, supra note 77 (“I wonder
if we would have been able to start Yelp today given Google’s recent actions.”).
130. See Goldman, supra note 12, at 195–96.
131. Competition, GOOGLE U.S. PUBLIC POLICY, http://www.google.com/
publicpolicy/issues/competition.html (last visited Nov. 27, 2013); Schmidt Tes-
timony, supra note 16, at 6; Adam Kovacevich, Google’s Approach to Competi-
tion, GOOGLE PUB. POLICY BLOG (May 8, 2009, 9:24 AM), http://
googlepublicpolicy.blogspot.com/2009/05/googles-approach-to-competition.html.
132. See, e.g., Volokh & Falk, supra note 13, at 884.
133. See, e.g., id. at 893–94.
134. See, e.g., Schmidt Testimony, supra note 16, at 132 (responding to
Sen. Grassley); Ammori & Pelican, supra note 101, at 11.
135. See, e.g., Goldman, supra note 12, at 197.
136. See, e.g., Goldman, supra note 86, at 99–100.
137. See, e.g., Schmidt Testimony, supra note 16, at 3–4.
138. See, e.g., Brief for Defendant, Infederation Ltd. v. Google Inc., [2013]
EWHC (Ch) No. HC12A02489, [15.2] (Eng.) (“Google cannot be held responsi-
ble for Foundem’s choice of business model. Foundem, at its own risk, appears
to have developed a business model that depends on its appearing high in free
search results.”); Search King, Inc. v. Google Tech., Inc., No. CIV-02-1457-M,
2003 BL 1897, at *7 (W.D. Okla. Jan 13, 2003) (“SearchKing consciously ac-
cepted the risk of operating a business that is largely dependent on a factor
(PageRank) over which it admittedly has no control.”); Danny Sullivan, Pen-
guin’s Reminder: Google Doesn’t Owe You a Living, So Don’t Depend on It,
MARKETING LAND (May 1, 2012), http://marketingland.com/penguin-google
-doesnt-owe-you-a-living-10968; see also Grimmelmann, Some Skepticism, su-
2014] SPEECH ENGINES 893
III. THE ADVISOR THEORY
The conduit and editor theories are not wrong. They are
merely incomplete. Each has valuable insights about the na-
ture of search, insights unique to the vantage points they
adopt. The conduit theory looks at search through websites’
eyes. The editor theory looks at search through search engines’
eyes. But we also can and should ask what search would look
like through users’ eyes.
Section A introduces the idea that users turn to search en-
gines for advice to help them decide among websites. Section B
gives a normative account of why we should prefer this user-
centric take on search. Section C translates this high-level the-
ory into a pair of policy prescriptions. And Section D considers
some limits on the advisor theory.
A. SEARCH RESULTS AS ADVICE
Over half a century ago, Vannevar Bush described the vast
informational universe we now inhabit:
Thus far we seem to be worse off than before—for we can enormously
extend the record; yet even in its present bulk we can hardly consult
it. This is a much larger matter than merely the extraction of data for
the purposes of scientific research; it involves the entire process by
which man profits by his inheritance of acquired knowledge. The
prime action of use is selection, and here we are halting indeed. There
may be millions of fine thoughts, and the account of the experience on
which they are based, all encased within stone walls of acceptable ar-
chitectural form; but if the scholar can get at only one a week by dili-
gent search, his syntheses are not likely to keep up with the current
139
scene.
For centuries, the idea that there is simply too much in-
formation in the world has been a persistent source of anxie-
140
ty. What is new in the age of the Internet is the sheer scale of
the problem. In 2011, humanity created and stored nearly two
141
zettabytes. The web contains over a trillion different webpag-
pra note 64, at 447–48 (questioning websites’ entitlement to traffic from
search engines).
139. Vannevar Bush, As We May Think, ATLANTIC MONTHLY, July 1945, at
118.
140. For histories of attempts to grapple with the problem, see, for exam-
ples, ANN BLAIR, TOO MUCH TO KNOW: MANAGING SCHOLARLY INFORMATION
BEFORE THE MODERN AGE (2010); JAMES GLEICK, THE INFORMATION: A HIS-
TORY, A THEORY, A FLOOD (2012); ALEX WRIGHT, GLUT: MASTERING INFOR-
MATION THROUGH THE AGES (2007).
141. JOHN GANTZ & DAVID REINSEL, IDC IVIEW: EXTRACTING VALUE FROM
CHAOS 1 (2011).
894 MINNESOTA LAW REVIEW [98:868
142
es. The world has over two billion Internet users, every single
143
one of whom is a potential speaker. If you want to listen to
them all in this lifetime, you have less than one second each—
assuming you do not stop to sleep or eat. We live in Borges’s
144
Library of Babel. Information itself is a good: the world would
not be better off if there were far less of it. Rather, the problem
is that the ratio of information to our ability to make use of it
has grown beyond all proportion.
This is a matching problem; the billions of speakers and
billions of listeners in the world need ways to decide who
speaks to whom at any given moment. We can approach it in
two fundamentally different ways. One way would be to try to
identify the best information sources and make sure they can
be heard through the cacophony. Although they differ on the
details, this is the approach taken by the conduit and editor
theories. The conduit theory worries that valuable and deserv-
ing speakers will be drowned out unless they have search en-
gines’ help. The editor theory sets up search engines as experts
in identifying the best and most useful information. Both are
speaker-oriented: they try to solve the problem of noise by am-
plifying good speech.
The alternative is listener-oriented: we could try to em-
power users to identify for themselves the speech they wish to
hear. An engineer would say that you can improve the signal-
to-noise ratio by using either a more powerful transmitter or a
more sensitive receiver. From a listener-oriented perspective,
then, a search engine is a tool for choosing which websites to
listen to.
Indeed, out of all the ways that speakers and listeners can
find each other, search is the single most listener-directed. The
entire point of consulting a search engine is that the user speci-
fies her own interests—not someone else’s—in the search query
and receives results relating to those interests. A search engine
that responds to [apple macbook] and [occupy cleveland]
and [stupid cat tricks] with the same list of results has
failed of its essential purpose. And users bring a truly remark-
142. Jesse Alpert & Nissan Hajaj, We Knew the Web Was Big . . ., GOOGLE
OFFICIAL BLOG (July 25, 2008), http://googleblog.blogspot.com/2008/07/we
-knew-web-was-big.html.
143. INTERNET WORLD STATS, http://www.internetworldstats.com/stats
.htm (last updated June 30, 2012) (estimating 2,405,518,376 Internet users).
144. See KEVIN KELLY, OUT OF CONTROL 258–63 (1994) (discussing Library
of Babel as a search problem); Grimmelmann, supra note 79 (giving extended
metaphor of the Internet as Borges’s infinite Library).
2014] SPEECH ENGINES 895
able range of interests to search engines. Compare a hundred-
channel cable system, or even a million-volume research li-
brary, with the four hundred and fifty billion distinct search
145
queries that Google has answered.
The crucial technological feature is interactivity. Unlike a
radio dial or a telephone directory, a search engine is not pre-
sented to users as a static artifact. Instead, search results are
generated “on the fly,” in response to a user’s specific query, in
a matter of milliseconds. Having hired the search engine once
to carry out a search, the user may decide to hire it again to
perform a related one. She can refine her query by entering
modified or additional keywords, seeing how this changes the
results. And when she is satisfied with the search engine’s sug-
gestions, she goes off to a website or websites to attend to their
speech.
On this view, search results are advice: suggestions about
which websites the user should consult. Calling search engines
advice-givers synthesizes the insights of the conduit and editor
theories. The ultimate goal of search, as the conduit theory ex-
146
plains, is to connect websites and users. Search engines can
advance this goal, as the editor theory explains, by expressing
147
judgments about websites. It is only from the user’s point of
view that these two functions are not opposites but two sides of
the same coin. Search engines connect websites and users by
expressing judgments about websites.
This characterization suggests a third normative theory for
evaluating search: a search engine should be a helpful, trust-
148
worthy advisor. An ideal advisor would have several im-
portant characteristics. It would adopt the user’s agenda, ra-
ther than trying to persuade the user of its own. It would be
perfectly omniscient; if the sought-after information exists at
145. See Welcome to Under the Hood, GOOGLE INSIDE SEARCH (Jan. 28,
2011), http://www.google.hu/insidesearch/underthehood.html.
146. See supra Part II.A.
147. See supra Part II.B.
148. This view of “advice” puts significant emphasis on its personalization.
A query-driven search engine of the sort discussed in this Article customizes
its results based on the queries a user enters, and often based on other charac-
teristics of the user that it is able to determine. Although there is obviously a
continuum between personalized and nonpersonalized advice—think of a hor-
oscope, which is personalized to the minimal extent of giving the reader one of
twelve different canned scripts—the distinction is in general significant. See,
e.g., Lowe v. S.E.C., 472 U.S. 181, 183 (1985) (holding that publishers of
“nonpersonalized investment advice and commentary” are not regulated by
the Investment Advisors Act of 1940).
896 MINNESOTA LAW REVIEW [98:868
all, the advisor would know where that information is. The ad-
visor would work quickly and cheaply. And having identified
the information the user seeks, the advisor would step aside
and let the user make her own decisions about what to do with
it.
B. ACTIVE LISTENING
The advisor theory has two basic commitments. First, it
puts users’ interests first, rather than websites’ or search en-
gines’: the goal of search is to help users find what they seek.
And second, it defers to users’ choices in defining those inter-
ests: the goal of search is to help users find what they seek.
What is so attractive about a world that gives users this capa-
bility? A great deal. It is a world of active listeners who are ca-
pable of exercising autonomous self-directed control over their
149
information diets. They seek out the speech they wish to hear
and avoid the speech they wish to ignore. The world is better
off for it, because the shift to active listening advances the val-
ues we care about. It promotes autonomy, equality, diversity,
150
and efficiency.
First, and most importantly, putting search users first
promotes autonomy, “an individual’s capacity to author her
151
life,” on two levels. First, there is the ability to choose appro-
149. Cf. Joseph P. Liu, Copyright Law’s Theory of the Consumer, 44 B.C. L.
REV. 397, 406–11 (2003). Liu focuses on how listeners engage with speech once
they know about it and have some measure of access to it; our focus here is on
how listeners find that speech in the first place.
150. This Section attempts, so far as possible, to rest these arguments
purely on listeners’ interests as listeners, rather than on their interests as fu-
ture speakers. Like Laplace, it has no need of that hypothesis.
151. Yochai Benkler, Siren Songs and Amish Children: Autonomy, Infor-
mation, and Law, 76 N.Y.U. L. REV. 23, 35 (2001). Autonomy is a contested
concept with multiple overlapping meanings. The sense in which it is used
here is based primarily on Benkler, at 32–41 and JOSEPH RAZ, THE MORALITY
OF FREEDOM 368–429 (1986). “An autonomous person’s well-being consists in
the successful pursuits of self-chosen goals.” RAZ, supra, at 370. Three features
of this definition are significant. First, autonomy in this sense is a theory of
“well-being”: a person can be more or less autonomous by degrees, and a per-
son is better off the more autonomous she is. See Benkler, supra, at 33–34. Se-
cond, the person’s goals are “self-chosen”: this form of personal autonomy
should be contrasted to the Kantian ideal of moral autonomy, in which auton-
omy consists in the rationally self-chosen adoption of universal moral laws.
See RAZ, supra, at 370 n.2. And third, this conception of autonomy requires
“successful” pursuit of one’s goals; it is therefore a theory of capabilities and
not just of authenticity.
This section deals only with a specific subset of obstacles to autonomy.
One’s autonomy can fail because of either “internal” or “external” constraints.
2014] SPEECH ENGINES 897
priate actions for achieving one’s goals. A farmer cares about
[sorghum yield improvement]; a questioning teen about
[ways to tell if your gay]. Both of them must make cru-
cial decisions. For them, as for all of us, knowing enough about
their options and their likely consequences can make the dif-
ference between success and failure, fulfillment and misery.
They have, in other words, a powerful autonomy interest as lis-
152
teners.
But very little of the information we want and need comes
to us of its own accord: we must go in search of it. The right to
“seek” information is so fundamental it is recognized in the
Universal Declaration of Human Rights and the International
153
Covenant on Civil and Political Rights. To be effective, this
active process of seeking out information must include not just
mere access to raw information, but also the ability to sort
through it. Just as it is not sufficient to put the farmer and the
teen in a field and shout at them the things we think they want
Benkler, supra, at 35–38. Internal constraints include the lack of “capacity” to
make plans for achieving one’s goals and the lack of “will” to stick to those
plans once made. Id. at 35–36. External constraints, on the other hand, inter-
fere either with one’s ability to choose one’s own goals or with one’s ability to
take actions that advance those goals. See id. at 36–38. And one’s ability to
take effective action can fail either because of a lack of options or a lack of
knowledge about them. RAZ, supra, at 377–78. Out of these five kinds of obsta-
cles—the lack of capacity, will, choice of goals, options, or knowledge—this sec-
tion considers two: interference with one’s choice of goals, and interference
with one’s knowledge of one’s options for achieving those goals. In Raz’s ter-
minology, the section deals with “manipulation” and not with “coercion.” Id. at
373.
152. For listener autonomy justifications of free speech protections, see,
e.g., Marc Jonathan Blitz, Constitutional Safeguards for Silent Experiments in
Living: Libraries, The Right to Read, and a First Amendment Theory for an
Unaccompanied Rights to Receive Information, 74 UMKC L. REV. 799 (2006);
Caroline Mala Corbin, The First Amendment Right Against Compelled Listen-
ing, 89 B.U. L. REV. 939, 953, 965 (2009); Richard H. Fallon Jr., Two Senses of
Autonomy, 46 STAN. L. REV. 875, 878 (1994); David A.J. Richards, Free Speech
and Obscenity Law: Toward a Moral Theory of the First Amendment, 123 U.
PA. L. REV. 45, 62 (1974); Thomas Scanlon, A Theory of Freedom of Expression,
1 PHIL. & PUB. AFF. 204, 209 (1972); David A. Strauss, Persuasion, Autonomy,
and Freedom of Expression, 91 COLUM. L. REV. 334, 355 (1991); Eugene
Volokh, Freedom of Speech in Cyberspace from the Listener’s Perspective: Pri-
vate Speech Restrictions, Libel, State Action, Harassment, and Sex, 1996 U.
CHI. LEGAL F. 377.
153. Universal Declaration of Human Rights, G.A. Res. 217(III)A, ¶ 19,
U.N. Doc. A/RES/217(III) (Dec. 10, 1948); International Covenant on Civil and
Political Rights, G.A. Res. 2200(XXI)A, ¶ 19(2), at 52, U.N. Doc. A/6316 (Dec.
19, 1966). See generally Molly Land, Toward an International Law of the In-
ternet, 54 HARV. INT’L L.J. 393 (2013) (discussing the history and interpreta-
tion of these provisions of the UDHR and ICCPR).
898 MINNESOTA LAW REVIEW [98:868
to learn, it is not sufficient to sit them down at a keyboard and
tell them to have at it. “Selection,” as Vannevar Bush called it,
154
is the missing link.
In particular, empowering users with search protects their
155
autonomy from manipulation. The editor theory depends on
search engines to know what is best for users; the conduit theo-
ry depends on websites to do the same. But users themselves
are better placed to know what they want and need than any-
one else is. A false claim to have the “lowest propane prices in
town!” only works on those who can’t search for [lowest pro-
pane prices in town]; a woman seeking [abortion in-
formation] will be offered more useful information and a wid-
er range of options if she uses a search engine than if she calls
the number on a “Pregnant? Need Help?” billboard.
Access to information, of which the ability to search is a
crucial component, also matters to autonomy for a second rea-
son. Our goals themselves—not just the means for pursuing
156
them—must be self-chosen for us to be truly autonomous. We
require what Michael Zimmer calls “intellectual mobility,”
which he defines as “the freedom to learn new things, explore
new ideas, adapt, and change one’s thoughts and beliefs in or-
157
der to grow and develop intellectually as an individual.” The
farmer’s goals—grow more sorghum—seem stable for now, but
the teen is engaged in the painful process of determining (not
158
merely discovering) who he is. He will emerge with a new-
found sense of how he wishes to live his life, one he could not
have specified in advance. If his informational experience lacks
queer voices because his search tools are unable to find them,
the self-determination at the core of his autonomy is threat-
159
ened.
154. Bush, supra note 128, at 121.
155. See RAZ, supra note 151, at 377–78 (discussing coercion and manipula-
tion as constraints on independence); Benkler, supra note 151, at 38–49 (dis-
cussing role of information environment in limiting or promoting autonomy).
156. See Benkler, supra note 151, at 55. See generally GERALD DWORKIN,
THE THEORY AND PRACTICE OF AUTONOMY (1988) (discussing preference-
formation aspects of autonomy).
157. Michael Zimmer, The Quest for the Perfect Search Engine: Values,
Technical Design, and the Flow of Personal Information in Spheres of Mobility
(unpublished PhD dissertation, NYU 2007).
158. See JULIE E. COHEN, CONFIGURING THE NETWORKED SELF: LAW,
CODE, AND THE PLAY OF EVERYDAY PRACTICE 84 (2012) (discussing the “emer-
gent self”).
159. Cf. Parents, Families, and Friends of Lesbians and Gays, Inc. v. Cam-
denton R-III Sch. Dist., 853 F. Supp. 2d 888, 892 (W.D. Mo. 2012) (enjoining
2014] SPEECH ENGINES 899
A second major virtue of widespread access to search is in-
formational equality. First, there is equality among users: egal-
160
itarian access to knowledge requires something like search. If
Affluent Amy has a personal lactation consultant on retainer
while Backwoods Barbara is fifteen miles from the nearest doc-
tor, it goes a long way toward making up the difference if they
both can search for [is it safe to breastfeed on su-
161
dafed]. Then, there is equality between listeners and speak-
ers. The capacity to listen is distributed far more evenly than
162
the capacity to speak. There are billionaires, but no one has a
billion ears. Disparities in wealth drop away when matching is
controlled by user interest rather than by who can flood the
airwaves with the most pervasive advertising.
Third, equality of access plus individual autonomy equals
diversity. It is precisely because people have wildly diverging
needs, capabilities, values, preferences, worldviews, and life
experiences that the individuation of search matters. A parent
worrying about [minor child bail eligibility] has vast-
ly different informational needs than a recent arrival in town
looking for [thai groceries in fresno]. A fifth of the que-
ries Google sees each day are new: no one else has ever used the
163
same combination of terms. The development of personalized
and social search is not just a means towards “better” results, it
is also a way of accommodating diversity of user interests. One
man’s noise is another man’s signal; delegating to users the de-
cision of what to search for lets them make different decisions.
Search also promotes diversity on the level of groups rather
school district from using Internet filter that “systematically allows access to
websites expressing a negative viewpoint toward LGBT individuals by catego-
rizing them as ‘religion,’ but filters out positive viewpoints to-
ward LGBT issues by categorizing them as ‘sexuality’”); see also Benkler, su-
pra note 151, at 42–45 (“Children who grow up with knowledge of a wide
variety of options for living, and with the capacity to value different options,
will have a greater role in selecting the option they will pursue as adults than
children who know no other option.”).
160. On access to knowledge and distributive values, see Lea Shaver, The
Right to Science and Culture, 2010 WIS. L. REV. 121.
161. The answer is yes, Sudafed is safe for the baby—but nursing mothers
may want to avoid it anyway, as it can cause irritability and decrease milk
production. See Nat’l Lib. of Med., Pseudoephedrine, LACTMED: DRUGS AND
LACTATION DATABASE, http://toxnet.nlm.nih.gov/cgibin/sis/search/f?./temp/~
HQ6lW0:1 (last updated Sept. 17, 2013).
162. See Eugene Volokh, Cheap Speech and What It Will Do, 104 YALE L.J.
1805, 1834–41 (1995) (discussing comparative empowerment of listeners in
digital environment).
163. Welcome to Under the Hood, supra note 145.
900 MINNESOTA LAW REVIEW [98:868
than individuals: it facilitates the development of minority and
micro-minority viewpoints like [dont drone me bro],
[baha’i homeschooling], and [pinealoma support
group], because it helps people with shared interests find each
other.
Finally, putting good search in users’ hands advances the
efficiency values already embedded in the numerous bodies of
164
law that lay claim to regulate the search process. Frequently
these bodies’ own normative frameworks start from the per-
spective of consumers at large—that is, from the perspective of
165
search users. So, for example, copyright law is designed to
166
“advance public welfare through the talents of authors” by of-
fering a “reward to the author or artist [that] serves to induce
167
release to the public of the products of his creative genius.”
Good search ensures that authors and publishers actually face
the incentive that copyright wants them to face: public demand
for their work. Trademark is designed to minimize consumer
search costs—that is, to make users maximally effective at
168
finding the goods and services they seek. Antitrust law focus-
169
es on consumer welfare. And so on. In each case, adopting us-
ers’ point of view aligns our understanding of search with our
other goals for search policy.
This is admittedly an idealized portrait of search. The
search engines we have today deviate from it in many respects.
The discussion of the conduit and editor theories above illus-
trates some of the numerous ways in which search engines fall
short, and we shall see further examples below. Search engines
are less than fully helpful for many searches, and entirely un-
helpful for some. The consequences fall unevenly and not al-
ways visibly on users and websites. But for all that, user em-
powerment is still a worthy ideal, and the closer we can bring
search engines to it, the better off users will be. The question is
how.
164. See Grimmelmann, supra note 1, at 15–51 (detailing thirteen such
bodies of law).
165. Id. at 7, 16.
166. Mazer v. Stein, 347 U.S. 201, 219 (1954).
167. United States v. Paramount Pictures, 334 U.S. 131, 158 (1948).
168. William M. Landes & Richard A. Posner, Trademark Law: An Eco-
nomic Perspective, 30 J. L. & ECON. 265, 281 (1987).
169. See Nat’l Collegiate Athletic Ass’n v. Bd. of Regents of Univ. of Okla.,
468 U.S. 85, 107 (1984).
2014] SPEECH ENGINES 901
C. ACCESS AND LOYALTY
“The perfect search engine would be like the mind of God,”
capable of anticipating a user’s needs and satisfying them be-
170
fore the user can even think to ask. But like divinity, perfect
search is unattainable in this life. Instead, the legal system
must deal with the institutional framework of search as it is
and could be. Law typically responds in two ways when it
comes across an advice-giving relationship. On the one hand, as
discussed in Section 1, it tries to ensure that people have access
to advisors. And on the other, as discussed in Section 2, it tries
to ensure that people can expect loyalty from their advisors.
Both principles apply here: users need search engines, and they
need to be protected from search engines.
1. Access
Little more need be said about why access is an important
value for users. Instead, the question is what law can do to
promote it. The answer has both negative and positive dimen-
sions.
On the negative side, some kinds of regulation obviously
171
threaten access. German law prohibits Holocaust denial; Thai
172
law prohibits insulting the king. Google frequently removes
links to these and many other kinds of content when ordered to
173
do so by local authorities. These deletions directly inhibit us-
ers’ ability to seek out the information they seek. The German
government doesn’t let users make up their own minds about
the Holocaust; the Thai government doesn’t let them decide
whether its monarchy is worthy of respect. When the Chinese
search engine Baidu blocks searches for information on Falun
174
Gong at the Chinese government’s behest, it interferes not
170. Charles Ferguson, What’s Next for Google, TECH. REV., Jan., 2005, at
40, available at http://www.technologyreview.com/featuredstory/403532/whats
-next-for-google/page/2/ (quoting Sergey Brin, co-founder of Google).
171. STRAFGESETZBUCH [STGB] [PENAL CODE], Nov. 13, 1998,
BUNDESGESETZBLATT, Teil I [BGBL. I], as amended, § 130(3) (Ger.).
172. See Thailand Sentences Editor to 10 Years in Jail for Royal Insult,
ASIANCORRESPONDENT.COM (Jan. 23, 2013, 12:53 PM), http://www
.asiancorrespondent.com/96003/thailand-sentences-editor-to-10-years-in-jail
-for-royal-insult/.
173. See generally Transparency Report, GOOGLE, http://www.google.com/
transparencyreport/removals/government/?hl=en (last visited Nov. 27, 2013)
(summarizing requests from government agencies and courts, and providing
specific examples of Google’s response to such requests).
174. See, e.g., Ed Felten, Chinese Internet Censorship: See It for Yourself,
FREEDOM TO TINKER (Apr. 28, 2009), https://freedom-to-tinker.com/blog/felten/
902 MINNESOTA LAW REVIEW [98:868
just with users’ religious freedom, but with a basic precondition
of that freedom.
Even when the government stops short of deleting or dic-
tating search results, its regulations can still threaten access. If
search engines weren’t allowed to use location information out
of privacy concerns, they couldn’t direct users to local business-
es rather than ones halfway around the world. The same ap-
plies to any signal in search engines’ repertoire: limiting its use
potentially degrades the quality of advice users receive. Even
seemingly collateral regulations can inhibit access. Advertising
is the economic engine behind the modern search engine as we
175
know it; if keyword advertising were illegal, we wouldn’t have
Google or Bing. Access therefore can have a libertarian valence:
governmental regulation of search is problematic because it re-
stricts users’ ability to consult the search engines they might
have preferred.
Positively, access can also have a liberal valence: the gov-
ernment should take steps to ensure that users are affirmative-
ly able to make use of good and diverse search engines, helping
176
to provide them if the market falls short. Law has at its dis-
posal the usual tools of information policy: government subsi-
dies, effective competition policy, good technical and legal in-
frastructure, education, and so on. The choice among these
tools is a matter of praxis, context, and culture. There are many
roads to relevance. From the user’s point of view, it does not
matter whether relevant search results are provided by gov-
ernment-subsidized academic research, by a dominant incum-
bent with the resources to invest heavily in product develop-
ment, by a Schumpeterian succession of innovative search
paradigms, or by cutthroat price and feature competition
among multiple search engines. Which of these will work
best—and what technology and competition policy will best
promote it—is an empirical question. What is good for Google
might be good for users, or it might not.
chinese-internet-censorship-see-it-yourself/.
175. See infra note 323.
176. For a particularly ambitious, comprehensive, and inspiring statement
of this form of access, see SIVA VAIDHYANATHAN, THE GOOGLIZATION OF EVE-
RYTHING (AND WHY WE SHOULD WORRY) 204–10 (2011) (describing a proposed
“Human Knowledge Project” that would be “open, public, global, multilingual,
and focused”).
2014] SPEECH ENGINES 903
2. Loyalty
There is an inescapable information asymmetry between
users and search engines. No one setting out on a journey of en-
lightenment knows what lies at the end of the road—if she did,
there would be no need of the journey. The user knows more
about what she wants, whether it be [free online calculus
practice questions] or [brinty spiers topless], but
the search engine knows far more about whether anyone has
put practice questions online and where those Britney Spears
177
pictures are.
This creates a distinctive possibility for disloyalty. If I
search for [discount dingos] and the search engine tells me
about OtterWorld and CapybaraCentral but not DingoMart, it
has frustrated my dingo-related goals. Perhaps worse, if the
search engine directs me to DingoBarn because it earns an un-
disclosed 5% commission on referrals, it has abused my trust to
178
enrich itself. It is precisely because the search engine knows
more than I do about websites that it can hide what it knows
from me, or deliberately steer me to sites that serve its goals,
not mine. Economically, this is a principal-agent problem. I
cannot fully trust the search engine to exert itself fully on my
behalf, because I am not fully capable of monitoring it. The
asymmetry is hard-wired into search; it is not possible to imag-
179
ine the user-search engine relationship without it.
Thus, the government can help searchers by taking action
against search engines that deceive or manipulate, or coerce
users. Loyalty might, as the editor theory predicts, arise purely
from competition among search engines. But where loyalty does
not come about on its own, law can step in to ensure that it
does.
177. Cf. Julie E. Cohen, The Place of the User in Copyright Law, 74
FORDHAM L. REV. 347, 349 (2005) (developing theory of the “situated user”
whose “patterns of consumption and the extent and direction of her own au-
thorship” are incompletely formed when she engages with works, and are
“shaped and continually reshaped by the artifacts, conventions, and institu-
tions that make up her cultural environment”). Although active listening and
the advisor theory ascribe a greater degree of agency to users than Cohen
does, they share with her the idea that users are engaged in a process of self-
development.
178. Cf. 16 C.F.R. § 255.5 (2013) (requiring disclosure of any “connection
between the endorser and the seller of the advertised product which might
materially affect the weight or credibility of the endorsement”).
179. Cf. Mark R. Patterson, Non-Network Barriers to Network Neutrality,
78 FORDHAM L. REV. 2843, 2860 (2010) (discussing the user’s inability to know
when or whether search engines are returning biased results).
904 MINNESOTA LAW REVIEW [98:868
The body of law most clearly concerned with problems of
disloyalty is fiduciary law, which monitors trustees, guardians,
doctors, corporate directors, and others who “enjoy[] discretion-
ary power over the significant practical interests of another”
180
within a particular domain. The case for applying fiduciary
concepts to search engines has two parts: finding a fiduciary re-
lationship and specifying the fiduciary duties that relationship
entails.
Search engines are not on the list of traditional fiduciaries,
but the list is not closed. Some courts have recognized spouses
181
as fiduciaries for each other and we are undergoing some-
thing of an academic fiduciary renaissance, with scholars argu-
182 183 184
ing for treating legislators, judges, jurors, and even
185
friends as fiduciaries. The common themes of fiduciary rela-
186
tionships are dependence, trust, and vulnerability. The
search engine provides a valuable service from a position of su-
perior knowledge and superior skill; the user provides it with
valuable and often sensitive information, trusting in it to pro-
187
vide suggestions consistent with her interests. Search en-
188
gines resemble lawyers and investment advisors, both of
whom give advice to their clients and are regarded as fiduciar-
189
ies when they do.
A useful source for fleshing out the relevant fiduciary du-
190
ties is agency law. Search engines are probably not agents as
180. Paul B. Miller, A Theory of Fiduciary Liability, 56 MCGILL L.J. 235,
262 (2011).
181. E.g., Dunkin v. Dunkin, 986 P.2d 706, 711–12 (Or. Ct. App. 1999).
182. E.g., D. Theodore Rave, Politicians as Fiduciaries, 126 HARV. L. REV.
671 (2013).
183. Ethan J. Leib et al., A Fiduciary Theory of Judging, 101 CALIF. L.
REV. 699 (2013).
184. Ethan J. Leib et al., Fiduciary Principles and the Jury, 55 WM. &
MARY L. REV. (forthcoming 2014).
185. Ethan J. Leib, Friends as Fiduciaries, 86 WASH. U. L. REV. 665 (2009).
186. See, e.g., Tamar Frankel, Fiduciary Law, 71 CALIF. L. REV. 795, 800 &
n.17, 810 (1983).
187. See VAIDHYANATHAN, supra note 176, at 59.
188. See MODEL RULES OF PROF’L CONDUCT R. 1.1 (2012) (duty of compe-
tence); R. 1.3 (duty of diligence); R. 1.4 (duty of communication and informed
consent); R. 1.6 (duty of confidentiality).
189. See 15 U.S.C. § 80b-6 (2012) (regarding prohibited transactions by in-
vestment advisers).
190. Cf. Mark R. Patterson, On the Impossibility of Information Intermedi-
aries 1–2 (Fordham Univ. Sch. of Law, Law & Econ. Research Paper No. 13),
available at http://papers.ssrn.com/sol3/papers.cfm?abstract_id=276968 (de-
scribing information intermediaries as agents that facilitate relationships be-
2014] SPEECH ENGINES 905
such: an agent undertakes to “act on the principal’s behalf and
191
subject to the principal’s control” and a search engine does
not typically deal with others on the user’s behalf, nor does the
user have control over the indexing and ranking process. But
the portions of agency law that deal with an agent’s duties to
its principal are instructive. An agent owes a fundamental duty
“to act loyally for the principal’s benefit,” a duty encompassing
192
fidelity, care, confidentiality, and disclosure. Thus, we might
say that a search engine must not let its own conflicts of inter-
193
est shape the results it gives a user; must not deliberately
underplay its hand in returning results it knows not to be rele-
194
vant; must not misuse the sensitive search queries she sup-
195
plies it with; and must not conceal important facts about how
196
it generates search results. All of these duties can be waived
with the user’s consent, but that consent must be both informed
197
and obtained in good faith.
Transparency is a crucial aspect of loyalty. On one level,
proper disclosures can defuse almost any deception. But on a
deeper level, transparency is also profoundly helpful in ena-
bling users to understand what it is they are getting from a
search engine and how to use it effectively. Google’s Inside
Search blog, which posts discussion of algorithmic additions
and describes how Google goes about creating search results, is
tween providers and seekers of information).
191. RESTATEMENT (THIRD) OF AGENCY § 1.01 (2006).
192. See id. § 8.01 (discussing “duty to act loyally for the principal's benefit
in all matters connected with the agency relationship”).
193. See id. § 8.02 (discussing “duty not to acquire material benefit from a
third party in connection with transactions conducted or other actions taken
on behalf of the principal”); id. § 8.03 (discussing duty not to act adversely to
the principal).
194. See id. § 8.08 (“If an agent claims to possess special skills or
knowledge, the agent has a duty to the principal to act with the care, compe-
tence, and diligence normally exercised by agents with such skills or
knowledge.”).
195. See id. § 8.05(2) (discussing duty “not to use or communicate confiden-
tial information of the principal for the agent's own purposes or those of a
third party”).
196. See id. § 8.11 (“An agent has a duty to use reasonable effort to provide
the principal with facts that the agent knows, has reason to know, or should
know when . . . the facts are material to the agent's duties to the principal
. . . .”).
197. See id. § 8.06 (“Conduct by an agent . . . does not constitute a breach of
duty if the principal consents to the conduct, provided that in obtaining the
principal's consent, the agent acts in good faith, discloses all material facts . . .
and otherwise deals fairly with the principal . . . .”).
906 MINNESOTA LAW REVIEW [98:868
hardly full and complete transparency, but is certainly a signif-
198
icant start.
3. Access and Loyalty Compared
In one sense, loyalty is merely a component of access:
search should be faithful to users’ goals, just like it should be
fast, comprehensive, and inexpensive. But loyalty is also in
tension with access, because the possibility of disloyal search
implies that sometimes bad search can be worse than no search
199
at all. The legal interventions needed to ensure loyalty may
sometimes have the effect of foreclosing a technical or business
model of search. Access alone might have no problem with
CorruptConcierge.com, which takes bribes from restaurants to
boost their rankings and conveniently “forgets” to tell users—
200
but loyalty surely would.
Access and loyalty are not binding legal rules. They are,
rather, “midlevel principles” that mediate between the pluralist
normative commitments described above and the nitty-gritty of
201
particular controversies. No law is prohibited because it vio-
lates access or is mandatory because it ensures loyalty. Appeals
to access and loyalty help us think through the consequences of
search engine practices, and help us devise legal strategies to
push those practices in the direction of users’ interests.
D. LIMITS OF THE ADVISOR THEORY
So far we have treated search users like the wise child at
the Passover Seder, asking good questions and receiving mean-
202
ingful answers. But what if they are more like one of the oth-
198. See, e.g., Amit Singhal, An Update to Our Search Algorithms, GOOGLE
INSIDE SEARCH (Aug. 10, 2012, 10:30 AM), http://www.insidesearch.blogspot
.com/2012/08/an-update-to-our-search-algorithms.html.
199. Transparency poses particularly complicated challenges. It can ad-
vance access by teaching search literacy, but it can also inhibit access by al-
lowing search engine optimizers to degrade the quality of search rankings by
cheating their way to the top. See Grimmelmann, supra note 1, at 55–56.
200. Cf. Chi-Chu Tschang, The Squeeze at China’s Baidu, BLOOMBERG
BUSINESSWEEK, Dec. 30, 2008, http://www.businessweek.com/stories/2008-12
-30/the-squeeze-at-chinas-baidu.htm (alleging that Chinese search engine
Baidu directly retaliates against sites that refuse to buy sponsored links by
demoting them in its organic rankings).
201. See generally ROBERT P. MERGES, JUSTIFYING INTELLECTUAL PROP-
ERTY 139–58 (2011) (developing theory of midlevel principles for intellectual
property law).
202. See generally NEW AMERICAN HAGGADAH (Jonathan Safra Foer & Na-
than Englander eds. 2012).
2014] SPEECH ENGINES 907
er three children: wicked, foolish, or simple? If so, the advisor
theory’s basic commitments—putting users’ interests first and
deferring to their self-definitions of those interests—may fail to
hold. These are all serious concerns, as we shall shortly see; the
advisor theory alone cannot fully address them. But—and this
is crucial—neither can the conduit and editor theories. Indeed,
they have much less to say, and they systematically obscure the
issues at stake when we try to shape search to serve public ra-
ther than private values. The advisor theory is a good first-
order approximation, even if taking more systemic concerns in-
2 3
to account will add x , x , and other higher-order terms.
First, there is the wicked child, who searches for [how to
build an h-bomb], [downtown abbey download epi-
sodes free], [chelsea clinton sex tape], or [kill all
the jews]. These search results hurt people: invading their
privacy, infringing their copyrights, promoting violence against
them, and so on. There is a pattern here. The people who are
203
harmed by these results are bystanders. From the perspec-
tive of websites, users, and the search engine that connects
them, everything is going just fine. The conduit and editor the-
ories lead to precisely the same conclusion as the advisor theo-
ry: there is no problem here.
But that is precisely the issue from the victim’s perspec-
tive: for her, search works best when it works least. She has a
point. We have copyright law, defamation law, child pornogra-
phy law, privacy law, and other kinds of information-limiting
laws for good reasons. They already reflect a considered social
judgment that some listeners—users—should be denied access
to speech they would like to receive. So users have an interest
in consulting search engines to help find information only
where it is information of a sort they have a legitimate interest
in receiving.
Next, there is the foolish child, who searches for [guy hit
in balls], [dumbass video getting hurt], [epic fail
waterski], [chainsaw accident], a thousand other varia-
tions on the same theme—and nothing else. This user is using
search to seal herself off in a private informational bubble con-
204
taining only humiliation and mutilation. But society, the
search for truth, and self-government all depend on dialogue
203. See Grimmelmann, supra note 1, at 33–44 (describing these as “third
parties’ interests” in search).
204. See ELI PARISER, THE FILTER BUBBLE: WHAT THE INTERNET IS HIDING
FROM YOU (2011); CASS R. SUNSTEIN, REPUBLIC.COM 2.0 (2007).
908 MINNESOTA LAW REVIEW [98:868
and civic education. The user who searches for [george bush
evil] or [climate change is bs] needs to be gently educat-
ed about different viewpoints; the user who searches for
[amercian idol] needs to be gently educated, period.
The conduit theory and the editor theory, being speaker-
oriented, cannot even rightly apprehend the nature of this ob-
jection. Bubbles trap listeners, not speakers. In contrast, the
bubble argument shares a central premise with the advisor
theory: users don’t know what information is available when
they seek it out. They also share a central goal: making users
better-informed. The difference is that where the advisor theo-
ry sees each search as a step toward enlightenment, the bubble
theory sees users as trapped in a cycle of self-reinforcing igno-
rance.
This is a difficult and much-debated subject, and a full dis-
cussion will need to await future work. For the moment, I will
offer four brief arguments that the bubble theory is a particu-
larly poor fit for search engines. First, it stands in stark con-
tradiction to one of the other most common complaints about
Google: that its ranking algorithms are too majoritarian, rather
than too individualized. It can’t both be the case that search
users are all looking at the same ten websites and that they’re
all living in their own individual information bubbles. Second,
there is a deep and abiding human taste for novelty, for seren-
dipity, for the unfamiliar. Search engines are particularly use-
205
ful in helping people explore new interests quickly and easily.
Third, the consequences of forcing search users to look at re-
sults they didn’t ask for and don’t want to see are dreadful. It
turns users into Alex from A Clockwork Orange, forcibly sub-
206
jected to high culture and unpleasant truths. This is a partic-
ular tragedy on the advisor theory, since the very point of
205. Cf. A. Lev-On, The Democratizing Effects of Search Engine Use: On
Chance Exposures and Organizational Hubs, in WEB SEARCH: MULTIDISCIPLI-
NARY PERSPECTIVES 135, 138–41 (Amanda Spink & Michael Zimmer eds.,
2008) (arguing that search engines “drive people to diverse and even opposing
views”). Eric Goldman argues that “Coasean filters [his term for good match-
ing systems] would not extinguish serendipitous exposures to unrequested
content because they would proactively generate content catering to consum-
ers’ latent interests.” Eric Goldman, A Coasean Analysis of Marketing, 2006
WIS. L. REV. 1151, 1219. He is right that the human taste for serendipity is
one that search engines and other matching systems do and should cater to,
but it is not really plausible to describe these tastes as “latent.” Pleasure and
knowledge are not just the surfacing of buried memories. Cf. Plato, Meno, in
FIVE DIALOGUES 58 (G.M.A. Grube trans., John M. Cooper 2d ed. 2002).
206. A CLOCKWORK ORANGE (Warner Bros. 1971).
2014] SPEECH ENGINES 909
search is that it can do so much more to enhance individual au-
tonomy and personal development. And fourth, if the fear is
that a personalized search engine will wrongly extrapolate from
a small sample of queries to trap the user in a bubble that dis-
torts her preferences, the problem is disloyal search engines,
not foolish users. Rather than being too user-directed, the
207
speech environment is not user-directed enough.
Finally, there is the simple child, who misunderstands
search results. She searches for [vaccination] and treats
Natural News (“Secret government documents reveal vaccines
208
to be a total hoax”) as authoritative because it was on Google;
or for [42 inch visio tv] and assumes the first result must
have the cheapest price because it is first; or for [obama mus-
lim] and doesn’t scroll down far enough to find the Wikipedia
209
entry. Studies have found that users trust search engines,
but also that they have woefully poor understandings of how
search engines work. The combination is dangerous, because it
causes overreliance on search results. Instead of independently
evaluating websites for themselves, users invest them with the
search engine’s authority.
Troublingly, Google shows every sign of wanting to push
even further. As Eric Schmidt put it, Google wants to be able to
answer questions like “What shall I do tomorrow?” or “What job
210
shall I take?” This is an autonomy-reducing relationship:
even if the search engine is capable of satisfying users, it is no
211
longer really helping them lead self-directed lives. When a
search engine usurps the user’s core decision-making authority,
it is hard not to describe the result as a serious violation of loy-
207. One useful policy intervention might be to require that search engines
must offer a non-personalized mode: a user must be able, at any time, to step
outside her bubble by disabling the customized filters the search engine has
created for her, and to receive generic, non-personalized results. Put another
way, the best remedy for bad search is more search.
208. Ethan A. Huff, Secret Government Documents Reveal Vaccines to Be a
Total Hoax, NATURALNEWS.COM (Jan. 8, 2013), http://www.naturalnews.com/
038598_vaccines_medical_hoax_government_documents.html.
209. See Pan et al., supra note 126.
210. Caroline Daniel & Maija Palmer, Google’s Goal: to Organise Your Dai-
ly Life, FIN. TIMES (London) (May 22, 2007, 9:08 PM), http://www.ft.com/
intl/cms/s/2/c3e49548-088e-11dc-b11e-000b5df10621.html#axzz2ggj5tsvs.
211. The best articulations of this fear come from science-fiction writers.
See Ken Liu, The Perfect Match, LIGHTSPEED (Dec. 2012), http://www
.lightspeedmagazine.com/fiction/the-perfect-match/; Tom Slee, Mr. Google’s
Guidebook, WHIMSLEY (Mar. 7, 2008), http://whimsley.typepad.com/whimsley/
2008/03/mr-googles-guid.html.
910 MINNESOTA LAW REVIEW [98:868
alty. Advice becomes a command; relevance gives way to some-
thing far more sinister. It is precisely for this reason that que-
ry-driven web search is better for autonomy than implicit rec-
ommendation systems like Facebook’s selections of which
212
stories from your friends to show you.
A final concern is the growing importance of distributed,
interactive, algorithmic processes in the sociotechnical co-
construction of meaning and authority. Choices made by pro-
grammers, publishers, and users feed back into each other re-
cursively with emergent, systemic consequences. At present, we
barely have the vocabulary to describe these processes, let
213
alone the theoretical frameworks to explicate them. They are
characterized by structures of information aggregation and dis-
tribution that are not necessarily intended or even compre-
hended by any of the contributors to those structures.
To the extent search engine law attempts to incorporate a
more systemic perspective, neither the conduit theory nor the
editor theory is much help. Both of them obscure the problem of
algorithmic authority. The conduit theory upholds an impossi-
ble ideal of neutrality; it can tolerate algorithms only to the ex-
tent that it fully specifies their results, that is, not at all. The
editor theory, by contrast, accepts whatever results from the
marketplace’s clash of algorithmic titans as an optimal out-
come. The editor theory systematically refuses to look inside
the algorithmic black box; the conduit theory smashes all such
boxes to smithereens. The advisor theory, by contrast, accepts
that we increasingly live in a world of algorithms and asks how
well they serve the goals of their users. It offers no special in-
sight into the workings of those algorithms, but it is prepared
to engage with those insights when other theories offer them
up. It is, at least, a place to start.
212. See Thomas E. Weber, Cracking the Facebook Code, THE DAILY BEAST
(Oct. 18, 2010, 2:57 AM), http://www.thedailybeast.com/articles/2010/10/18/
the-facebook-news-feed-how-it-works-the-10-biggest-secrets.html. Thus, there
are strong reasons to reject the convergence of search, advertising, and rec-
ommendation systems hailed by some commentators. E.g., Hector Garcia-
Molina et al., Information Seeking: Convergence of Search, Recommendations,
and Advertising, 54 COMM. ACM 121, 126 (2011). They may be technically sim-
ilar, but they are not normatively equivalent: one of them (search) is better
than the others.
213. A particularly useful survey of existing work and outline for future
efforts is Tarleton Gillespie, The Relevance of Algorithms, in MEDIA TECHNOL-
OGIES (Tarleton Gillespie et al. eds., forthcoming Jan. 2014).
2014] SPEECH ENGINES 911
IV. SEARCH BIAS RECONSIDERED
We are now in a position to apply the advisor theory to
specific legal problems. Search bias, the most controversial of
the many controversial claims against Google, is a good place to
start.
An early search bias lawsuit, Search King v. Google, is still
214
the leading example of the genre. Search King alleged that
Google reduced its PageRank—one of the most important sig-
nals used by Google to estimate a webpage’s importance—from
8 to 4, causing a precipitous drop in its traffic from Google and
215
a concomitant fall-off in business. Search King sued for tor-
tious interference with contractual relations, arguing that
Google devalued it “after and because Google learned that
216
[Search King] was competing with Google.”
Search King also shows what tends to happen to search-
bias claims in court: they lose. Search King claimed that
PageRanks were “objectively verifiable” and that Google
changed them “purposefully and maliciously,” rendering its
217
conduct “wrongful” and harming Search King. But Google re-
sponded, and the court agreed, that ranking decisions were
“fundamentally subjective,” so that there was “no conceivable
way to prove that the relative significance assigned to a given
218
web site is false.” As a consequence, Google’s search results
were “constitutionally protected opinions,” rendering them
219
“immune from tort liability.” The conduit theory met the edi-
tor theory, and the editor theory won. But both approaches are
too categorical. Search results are a mix of the objective and the
subjective. It is not possible to classify them as exclusively one
or the other.
214. No. CIV-02-1457-M, 2003 WL 21464568 (W.D. Okla. May 27, 2003).
For other notable search bias cases, see supra note 16. This Part will restrict
its attention to tort suits for misranking, rather than considering the whole
range of possible regulations to prevent search bias. The application of the
First Amendment to a tortious interference claim raises all of the essential is-
sues. And the approach this Part endorses, in which the search engine’s im-
munity turns on its good faith in answering users’ queries, is broadly applica-
ble to search bias issues, regardless of what doctrinal box they arrive packaged
in.
215. Search King, 2003 WL 21464568, at *1.
216. Id. at *2.
217. Id. at *2–3.
218. Id. at *3–4.
219. Id.
912 MINNESOTA LAW REVIEW [98:868
Instead, a better approach to search bias is to look at rank-
ings from users’ point of view. When a search engine gives ad-
vice to users, it speaks; there is no way to understand the giv-
ing of advice without implicating speech’s communicative
220
function. Moreover, a search engine’s advice is socially valua-
ble speech; we have seen an abundance of reasons why users as
listeners would suffer if this speech could be suppressed. But it
does not follow that search results ought to be categorically pro-
tected by the First Amendment. Precisely because they are val-
uable instrumentally rather than expressively, search results
should not be protected where they deceive the users they are
meant to inform.
This Part works through the proper First Amendment
221
analysis of a Search King-style tortious interference claim.
Section A unpacks search rankings to show that they consist of
the search engine’s opinions about relevance to the user. Sec-
tion B argues that the crucial doctrinal question in such a claim
is whether a ranking is demonstrably false and made with
knowledge of its falsity. Section C puts these two points togeth-
er. It argues that it is not possible to prove a ranking false by
proving it objectively wrong, but it is possible to prove a rank-
ing false by proving it subjectively dishonest. Or, more concise-
ly, Section A presents the facts, Section B the law, and Section
C the application of law to fact.
220. See Stuart Minor Benjamin, Algorithms and Speech, 161 U. PA. L.
REV. 1445 (2013). Another commentator, Andrew Tutt, argues, “Software, in
other words, should be considered not for what it is or even what it says but
for what it means to society to treat it like speech. Whether operating systems,
search engines, and word processors are ‘speech’ depends on the position these
categories occupy within our democracy.” Andrew Tutt, Essay, Software
Speech, 65 STAN. L. REV. ONLINE 73, 77 (2012). On the level of principles, he is
obviously right. But the only way to understand “the position these categories
occupy within our democracy” is to examine them closely, asking what mes-
sages software systems communicate to whom and how the listeners under-
stand those messages and act on them. This factually focused analysis
achieves Tutt’s goal of getting beyond unhelpful categories like “word proces-
sor” and “video game,” id. at 76, by bringing out the cultural and social prac-
tices associated with specific pieces of software in specific contexts.
221. This Part considers only tort claims, not potential search neutrality
regulations. With the end of the FTC’s search-bias investigation, no such regu-
lations appear likely in any jurisdiction where the First Amendment applies.
Any extended discussion of them would be painfully hypothetical, filled with
assumptions and subjunctives. That said, there is a strong argument that the
First Amendment would allow user-protective legislation and FTC enforce-
ment actions that go beyond the limits applicable to tortious-interference
claims. A few key points of this argument are set out in the margin infra notes
300 and 304.
2014] SPEECH ENGINES 913
The rest of the Part cleans up some loose threads. Section
D argues that it makes no difference that search results are
generated using computer algorithms rather than by hand. And
Section E pulls these claims together to defend, for the most
part, the Federal Trade Commission’s handling of the search
bias claims against Google.
A. SEARCH RANKINGS ARE OPINIONS ABOUT RELEVANCE
If there is one thing everyone can agree on in the search
bias debate, it is that the sine qua non of search results is rele-
vance. Compare Google critic Adam Raff’s demand that search
222
results should be “based solely on relevance” with Google en-
gineer Amit Singhal’s insistence that Google’s “algorithms rank
results based only on what the most relevant answers are for
223
users.”
Neither the conduit theory nor the advisor theory can pro-
vide a usable definition of relevance, because neither websites
nor search engines are proper judges of relevance. For every
website that gains in the rankings there is another that falls;
there is no way to break the tie between DingoMart and
DingoBarn’s competing claims. But if “relevance” is a quality
created by a search engine then it is devoid of meaning; the top
result for [dingo] is the most relevant, regardless of whether it
224
has anything to do with dingos. The conduit theory is inde-
terminate; the editor theory is tautological.
The advisor theory does better at capturing relevance; it
asks us to take users’ point of view. Through a user’s eyes, a
relevant result is one that helps her achieve her personal in-
formational goals; an irrelevant result is one that does not. As
one textbook explains:
A human is not a device that reliably reports a gold standard
judgment of relevance of a document to a query. Rather, humans and
their relevance judgments are quite idiosyncratic and variable. But
this is not a problem to be solved: In the final analysis, the success of
an [information retrieval] system depends on how good it is at satisfy-
222. Raff, supra note 2.
223. Amit Singhal, Setting the Record Straight: Competition in Search,
GOOGLE PUB. POL’Y BLOG (June 8, 2012), http://googlepublicpolicy.blogspot
.com/2012/06/setting-record-straight-competition-in.html.
224. Cf. James Grimmelmann, Three Theories of Copyright in Ratings, 14
VAND. J. ENT. & TECH. L. 851, 876 (2012) (critiquing claim that expressive rat-
ings are valuable simply because they are expressive).
914 MINNESOTA LAW REVIEW [98:868
ing the needs of these idiosyncratic humans, one information need at
225
a time.
Search engines face two linked problems in trying to satis-
fy users’ standards of relevance: the diversity of users with dif-
ferent intentions, and the difficulty of inferring intention from
a bare search query. “[E]ven if two people use the exact same
words to ask a question, they may be asking very different
226
things.” Google asserts that information about the Founding
Farmers restaurant is the most relevant result for [founding
farmers]. Perhaps it is, and many people would agree. But
what about a user looking for the February 2012 blog post from
the Paris Review Daily reviewing a modern edition of Martha
Washington’s family recipe collection—a post entitled “The
227
Founding Farmers”? For some users, this post is more rele-
vant than the restaurant’s homepage. Other users may be look-
ing for critical reviews of the restaurant or for amusing stories
228
about its knowledgeable but inattentive servers.
Search engines respond to the ambiguities of relevance in
three stages: they measure users’ satisfaction with search re-
225. CHRISTOPHER D. MANNING ET AL., INTRODUCTION TO INFORMATION
RETRIEVAL 151 (2008); see also KEN HILLIS ET AL., GOOGLE AND THE CULTURE
OF SEARCH 53–76 (2013) (discussing Google’s understanding of “relevance”);
Van Couvering, supra note 116 (“‘Really, it is the standard definition, which is,
we are trying to answer people's questions. Period. Relevance is when we ac-
tually return something that answers their question.’” (quoting “Interviewee
E”)).
226. @CopyrightLibn [Nancy Sims], TWITTER (July 11, 2013, 11:57 AM),
https://twitter.com/copyrightlibn/status/355400591998058497.
227. Robin Bellinger, The Founding Farmers, PARIS REV. DAILY (Feb. 23,
2012), http://www.theparisreview.org/blog/2012/02/23/the-founding-farmers/.
228. For examples of the ambiguities inherent in trying to satisfy users’
diverse standards of relevance, it is hard to beat the guidelines Google gives to
its human reviewers. See Search Quality Rating Guidelines 1.0, GOOGLE (Nov.
2, 2012), http://static.googleusercontent.com/external_content/untrusted_dlcp/
www.google.com/en/us/intl/en_us/insidesearch/howsearchworks/assets/
searchqualityevaluatorguidelines.pdf. “Many queries have more than one
meaning,” and raters are asked to classify them as “dominant,” “common,” or
“minor.” Id. § 2.3, at 8. For example, “The query [mercury], English (US)
might refer to the car brand, the planet, or the chemical element (Hg). While
none of these is clearly dominant, all are common interpretations. Many or
some people might want results related to these interpretations.” Id. The que-
ry [Nikon digital camera] seems clearer at first glance, but the guidelines note
that “Some users may have decided to buy a Nikon (‘do’), but some may be re-
searching the Nikon brand (‘know’), and some may want to go to digital cam-
era pages on the Nikon website (‘go’).” Id. § 2.4.4, at 11. Search engineers,
trademark lawyers, and lexicographers all know that words and phrases are
capable of sustaining multiple unrelated meanings, depending on the context
in which they appear.
2014] SPEECH ENGINES 915
229
sults, they predict how users will react to other results using
230
general theories of relevance, and they implement those theo-
231
ries in the algorithms that respond to users’ queries. Each
stage introduces its own approximations. At the measurement
stage, no focus group or A/B test is ever comprehensive enough
to capture the preferences of every user in the world; even if it
could, users would still misreport their long-term goals and
click on promising-looking results that turn out to be worthless
on further inspection. At the prediction stage, the search en-
gine must extrapolate from queries and webpages it has seen to
ones it has not. Extrapolations are guesses; guesses can be
wrong. And at the implementation stage, each algorithmic
tweak to improve relevance must be traded off against very real
costs. There are fixed costs, incurred simply to program and
test the tweak, and there are incremental costs, as each addi-
232
tional computation drives up power bills, hardware purchas-
233 234
es, and user waiting times. Different search engines use
different metrics, theories, and implementations, and hence
they deliver different search results.
Thus, search results are neither entirely “objective” nor en-
tirely “subjective.” The confusion that has surrounded the ques-
tion for a decade is a result of conflating users’ and search en-
gines’ views of relevance. From a user’s perspective, relevance
is a subjective goal. But from a search engine’s perspective,
search rankings are approximations of objectively but imper-
fectly observable characteristics of subjective user preferences,
embodied in the search engine’s choices about its algorithms. It
is these choices—disagreements about the most effective way to
229. See, e.g., id. at 6. Even this 43-page document is a “gutted” version of a
161-page set of guidelines in use internally. See Matt McGee, Google Gutted
Its Search Quality Rating Guidelines for Public Release, SEARCH ENGINE LAND
(Mar. 1, 2013, 4:08 PM), http://searchengineland.com/google-gutted-its
-search-quality-rating-guidelines-for-public-release-150281.
230. PageRank is the most obvious example: the theory behind using links
to rank pages is that users will tend to find more highly linked pages to be
more relevant. See Levy, supra note 23, at 98–99. The same is true for any
other signal a search engine uses: the signal is a theory about relevance. Id.
231. See id.
232. See Xiaobo Fan et al., Power Provisioning for a Warehouse-Sized Com-
puter, 34TH ACM CONF. ON COMPUTER ARCHITECTURE 13 (2007).
233. See Luiz André Barroso et al., Web Search for a Planet: The Google
Cluster Architecture, IEEE COMPUTER SOC’Y 22, 25 (2003).
234. See Jake Brutlag, Speed Matters for Google Web Search, GOOGLE
(June 22, 2009), http://services.google.com/fh/files/blogs/google_delayexp.pdf
(“Experiments demonstrate that increasing web search latency reduces the
daily number of searches per use by 0.2% to 0.6%.”).
916 MINNESOTA LAW REVIEW [98:868
measure and implement relevance—that constitute the “opin-
235
ions” in search.
There are, in other words, two different kinds of opinions
at work in search. Let us call them “normative” and “descrip-
236
tive,” respectively. Users have normative opinions about rele-
vance: expressions of the holder’s personal tastes and values.
But search engines have descriptive opinions about relevance:
claims about facts in the world under conditions of uncertainty.
Dale Peck’s assertion that Rick Moody is “the worst writer of
237
his generation” is a normative opinion; a forecaster’s predic-
tion that it will be 84º and sunny in Los Angeles tomorrow is a
238
descriptive opinion. Normative opinions are wholly subjec-
tive; descriptive opinions are subjective in that they express the
speaker’s personal belief about something not universally
agreed-on, but objective in that this something exists inde-
239
pendently of the speaker.
235. As this point should make clear, opinions about the best way to assess
relevance are not “speech” the First Amendment is concerned with. I may have
an opinion about the most effective way to mow my lawn; that does not mean
that mowing my lawn is speech. The resulting search rankings may be speech,
but that is because they communicate claims about relevance, not because
they communicate an idea about the best way to assess relevance.
236. This distinction is similar to one drawn in W. Page Keeton, Defama-
tion and Freedom of the Press, 54 TEX. L. REV. 1221, 1233–34 (1976) between
“evaluative” and “deductive” opinions. Keeton’s category of “deductive” opin-
ions, however, is inapt because he defines them in terms of how they are de-
rived—they are “drawn as an inference from the existence of other facts”—
rather than in terms of how they function as communicative acts—they make
claims that are in theory subject to observation by others. See also Wendy
Gerwick Couture, Opinions Actionable as Securities Fraud, 73 LA. L. REV. 382,
408–14 (2013) (extending Keeton’s distinction to securities law). A third kind
of “opinion” shows up in the case reports: “loose, figurative” language not
meant to be understood as making truth-valued claims at all. Old Dominion
Branch No. 496, Nat’l Ass’n of Letter Carriers v. Austin, 418 U.S. 264, 284
(1974); Diane Leenheer Zimmerman, Curbing the High Price of Loose Talk, 18
U.C. DAVIS L. REV. 359, 398–99 (1985) (identifying this third category of “opin-
ion”).
237. Dale Peck, The Moody Blues, NEW REPUBLIC, July 1, 2002, available
at http://www.newrepublic.com/article/books-and-arts/the-moody-blues (re-
viewing RICK MOODY, THE BLACK VEIL: A MEMOIR WITH DIGRESSIONS (2002)).
238. See generally L.A. STORY (TriStar Pictures 1991).
239. See Ronald K. Chen, Once More into the Breach: Fact Versus Opinion
Revisited After Milkovich v. Lorain Journal Co., 1 SETON HALL CONST. L.J.
331, 335 (1991) (describing what I call descriptive opinions as “speculation”);
Robert C. Post, The Constitutional Concept of Public Discourse: Outrageous
Opinion, Democratic Deliberation, and Hustler Magazine v. Falwell, 103
HARV. L. REV. 601, 657 (1990) (explaining that “verifiable” statements are
those for which “given enough time and effort, we would expect the claim to be
confirmed or disconfirmed by a consensus of investigators”). The underlying
2014] SPEECH ENGINES 917
In his Senate testimony in September 2011, Eric Schmidt
seemed to further confuse the objective/subjective debate when
240
he called Google’s rankings its “scientific opinion.” But this is
actually a helpful way of thinking about search results, per-
haps more so than Schmidt realized. Google studies the world,
draws conclusions, and shares them with the public, just as sci-
241
entists do. Google is not “scientific” in the sociological sense
that it publishes theories of relevance for peer evaluation. But
it aspires to be “scientific” in the sense of the Federal Rules of
Evidence’s definition of “scientific . . . knowledge,” which must
be “based on sufficient facts or data” and “the product of relia-
ble principles and methods” that are “reliably applied . . . to the
242
facts.” Scientific opinions are subjective to the extent that re-
ality is unknowable and scientists must forever make do with
dueling hypotheses and insufficient data. But they are objective
to the extent that they are based on reality and seek to describe
243
the world as it is. Search results seek to answer the imper-
244
fectly answerable question of what users want.
B. THE FIRST AMENDMENT REQUIRES FALSITY AND FAULT
Now that we have a handle on what search rankings are—
descriptive opinions—it is time to consider what First Amend-
ment standard applies to them. This Section argues that for a
assumption here, about the existence of “facts” in the world, passes over some
significant epistemological difficulties. For a more sophisticated treatment of
the issue, see Post, supra, at 656–61. For present purposes, nothing essential
depends on this precision.
240. Schmidt Testimony, supra note 16, at 234; cf., Dana Remus Irwin,
Freedom of Thought: The First Amendment and the Scientific Method, 2005
WIS. L. REV. 1479, 1481–92 (discussing scholarship on the application of the
First Amendment to scientific inquiry).
241. See Schmidt Testimony, supra note 16, at 234 (“constantly experi-
menting”).
242. FED. R. EVID. 702.
243. See ONY, Inc. v. Cornerstone Therapeutics, Inc., 720 F.3d 490, 496 (2d
Cir. 2013) (“Most conclusions contained in a scientific journal article are, in
principle, capable of verification or refutation by means of objective proof. In-
deed, it is the very premise of the scientific enterprise that it engages with
empirically verifiable facts about the universe. At the same time, however, it
is the essence of the scientific method that the conclusions of empirical re-
search are tentative and subject to revision, because they represent inferences
about the nature of reality based on the results of experimentation and obser-
vation.” (internal quotation marks omitted)).
244. See Our Products and Services, GOOGLE, http://www.google.com/about/
company/products/ (“Larry Page, our co-founder and CEO, once described the
‘perfect search engine’ as something that ‘understands exactly what you mean
and gives you back exactly what you want.’”).
918 MINNESOTA LAW REVIEW [98:868
statement of this sort to be actionable in a tortious interference
suit, it must be provably false and the speaker must have acted
with a sufficient degree of fault. (For present purposes, we can
assume that the fault standard is actual malice.) Subsection 1
locates this standard in the Supreme Court’s opinion in
Milkovich v. Lorain Journal Co.; Subsection 2 responds to con-
duit- and editor-theory objections that some other standard
ought to apply.
1. Milkovich
The First Amendment standard applied in Search King,
the leading search bias case against Google, comes from the
245
Supreme Court’s opinion in Milkovich v. Lorain Journal Co.,
a defamation case against a newspaper for a column saying
that “Anyone who attended the [wrestling] meet . . . knows in
his heart that [the plaintiffs] lied at the hearing after each hav-
246
ing given his solemn oath to tell the truth.” The newspaper
argued that the statement was protected as an “opinion,” but
the Supreme Court disagreed. There was no “wholesale defa-
mation exemption for anything that might be labeled ‘opinion,’”
and no need to divide statements into categories of “opinion” or
247
“fact.”
Instead, the Court applied its usual First Amendment pro-
tections against defamation. A statement about a public figure
on a matter of public concern by a media defendant “must be
provable as false before there can be liability” and the defend-
ant must have acted with actual malice, that is, “with
knowledge of their false implications or with reckless disregard
248
of their truth.” Thus, regardless of how they are labeled,
statements that knowingly “imply a false assertion of fact” can
249
be actionable.
245. Milkovich v. Lorain Journal Co., 497 U.S. 1 (1990).
246. Id. at 5.
247. Id. at 18.
248. Id. at 20–21.
249. It is still common to see in lower-court opinions a distinction between
actionable statements of “fact” and non-actionable statements of “opinion.”
These courts are misreading Milkovich, which rejected this very dichotomy.
See Robert D. Sack, Protection of Opinion Under the First Amendment: Reflec-
tions on Alfred Hill, “Defamation and Privacy Under the First Amendment,”
100 COLUM. L. REV. 294, 320–21 (2000). Sometimes, they distinguish “fact”
from “opinion” by saying that the former are capable of being proven false and
the latter are not. Since falsifiability was the basis of the Court’s holding in
Milkovich, no real harm is done. See Milkovich, 497 U.S. at 21. But when
courts draw on the vague contextual factors the Court rejected, the result is a
2014] SPEECH ENGINES 919
Milkovich thus directs us to ask two questions: does the
challenged speech make a false assertion of fact, and if so, is it
uttered with a sufficient degree of fault? The precise degree of
250
fault required may vary, but in thinking about search bias,
we can avoid these doctrinal difficulties by assuming that the
251
actual-malice standard applies. Search King, like other
search-bias plaintiffs, alleged that Google acted with the intent
to cause harm and with knowledge that its manipulated rank-
252
ings were false.
2. Some Objections
On this view, showings of falsity and fault are both neces-
sary and sufficient to hold a search engine liable in tort for
misranking a website. This position has been challenged from
both sides. Conduit theorists have said that falsity and fault
are not necessary; editor theorists have said that falsity and
fault are not sufficient. Both are wrong; Milkovich provides the
253
correct standard.
On the conduit side, Oren Bracha and Frank Pasquale
point to “the uncovered speech in an aircraft navigational
hopeless muddle. See, e.g., Gilbrook v. City of Westminster, 177 F.3d 839, 861
(9th Cir. 1999); Sack, supra, at 324–25 (criticizing Gilbrook and similar opin-
ions).
250. The Court has been less concerned, for example, about protecting
statements about private individuals or on matters of private concern. See
Gertz v. Robert Welch, Inc., 418 U.S. 323, 347 (1974) (“[S]o long as they do not
impose liability without fault, the States may define for themselves the appro-
priate standard of liability” for an award of actual damages for a false state-
ment by a media defendant about a private individual.); Dun & Bradstreet,
Inc. v. Greenmoss Builders, Inc., 472 U.S. 749, 760–61 (1985) (allowing an
award of presumed and punitive damages for a false statement on a matter of
private concern even without actual malice). The holding of Greenmoss is elu-
sive, as no opinion commanded a majority. See generally Lee Levine & Ste-
phen Wermiel, The Landmark that Wasn’t: A First Amendment Play in Five
Acts, 88 WASH. L. REV. 1 (2013) (reviewing in detail the Justices’ deliberations
and internal debates over Greenmoss).
251. See New York Times Co. v. Sullivan, 376 U.S. 254, 279–80 (1964).
252. See Search King, Inc. v. Google Tech., Inc., No. CIV-02-1457-M, 2003
WL 21464568, at *2–3 (W.D. Okla. May 27, 2003).
253. This claim requires one qualification, albeit one that is tangential to
the concerns at stake in the search-bias debate. After United States v. Alvarez,
132 S.Ct. 2537 (2012), it is likely that a prohibition on false speech must also
target a “subset of lies where specific harm is more likely to occur.” Id. at 2555
(Breyer, J., concurring). For present purposes, this is unlikely to be an issue.
Search bias suits like Search King sound in the well-established intentional
interference tort, which sufficiently limits suits to an appropriate “subset” of
cases in which the plaintiff can allege genuine economic injuries.
920 MINNESOTA LAW REVIEW [98:868
254
chart” in the case of Saloomey v. Jeppesen & Co., where the
defendant published a chart incorrectly stating that the Mar-
255
tinsburg, WV airport had an instrument landing system. An
experienced pilot using the chart tried to make an instrument
landing at Martinsburg and crashed on approach, killing him-
256
self, his father, and his son. The Second Circuit affirmed a ju-
257
ry verdict that the chart was a defective product. Saloomey
and other aircraft-chart cases generally do not even discuss the
First Amendment; Bracha and Pasquale would say that it is
258
similarly inapplicable to search rankings.
The result in Saloomey makes sense, but let us be clear
about the reason. The chart was defective because the infor-
mation it presented was false. If the chart had truthfully de-
scribed the facilities available at the Martinsburg airport, the
First Amendment would have barred the suit. Jeppesen has a
First Amendment right to sell accurate charts, but not to sell
259
inaccurate ones. In other words, the factual speech at issue in
Saloomey and the other products liability cases is not uncov-
ered speech, which is “measured against no First Amendment
standard whatsoever” and for which “[t]he First Amendment
260
just does not show up.” Rather, it is unprotected speech,
which faces “the full arsenal of First Amendment rules, princi-
ples, standards, distinctions, presumptions, tools, factors, and
261
three-part tests” and is found wanting.
On the editor side, Eugene Volokh and Donald Falk cite
Winter v. G.P. Putnam’s Sons for the proposition that search
262
results are pure expression and fully protected speech. There,
254. Bracha & Pasquale, supra note 56, at 1194 & n.237.
255. Saloomey v. Jeppesen & Co., 707 F.2d 671, 672 (2d Cir. 1983).
256. Id. at 672–73.
257. Id. at 672.
258. Another navigational-chart case of note is Brocklesby v. United States,
767 F.2d 1288, 1295 n.9 (9th Cir. 1985), which refused to consider a First
Amendment defense because the defendant raised the issue for the first time
on appeal. An earlier but withdrawn panel opinion rejected the defense, call-
ing the chart a “false or misleading commercial message[].” Brocklesby v.
United States, 753 F.2d 794, 803 (9th Cir. 1985), withdrawn and amended,
767 F.2d 1288 (9th Cir. 1985).
259. See, e.g., Sorrell v. IMS Health, Inc., 131 S. Ct. 2653, 2670–71 (2011)
(striking down law prohibiting sale of truthful information).
260. Frederick Schauer, The Boundaries of the First Amendment: A Prelim-
inary Exploration of Constitutional Salience, 117 HARV. L. REV. 1765, 1769
(2004).
261. Id.
262. See Volokh & Falk, supra note 13, at 890.
2014] SPEECH ENGINES 921
the plaintiffs were “mushroom enthusiasts who became severe-
ly ill from picking and eating mushrooms after relying on in-
formation in The Encyclopedia of Mushrooms, a book published
263
by the defendant.” As a matter of tort law subject to the “gen-
tle tug of the First Amendment and the values embodied there-
in,” the Ninth Circuit held that the publisher had no “duty to
investigate the accuracy of the contents of the books it publish-
264
es.” For Volokh and Falk, Google is a publisher just like Put-
265
nam: its speech is both covered and protected.
But Winter does not stand for the proposition that the First
Amendment absolutely shields the publishers of harmful misin-
formation. In a footnote, the court added that a “stronger ar-
gument might be made” in a case involving “fraudulent, inten-
266
tional, or malicious misrepresentation.” Winter, in other
words, is a case about fault rather than falsity: strict liability or
negligence will not support a lawsuit against a book publisher,
267
but worse misconduct may. That has to be right. The First
Amendment should not shield a publisher who advises readers
to eat Amanita phalloides with liver failure aforethought. For
prudential reasons, it makes sense to hold Jeppesen to a higher
standard of care than Putnam given the nature of their respec-
268
tive publications. But in both cases, the crucial threshold of
263. Winter v. G.P. Putnam’s Sons, 938 F.2d 1033 (9th Cir. 1991).
264. Id. at 1037.
265. See Volokh & Falk, supra note 13, at 892.
266. Winter, 938 F.2d at 1037 n.9.
267. The same principle applies to other lawsuits by readers who were in-
jured after following bad advice in publications. See, e.g., Cardozo v. True, 342
So. 2d 1053, 1054 (Fla. Dist. Ct. App. 1977) (cookbook with recipe including
taro roots, which are toxic when eaten raw); Smith v. Linn, 563 A.2d 123, 124
(Pa. Super Ct. 1989) (book recommending “liquid protein diet”); Herceg v. Hus-
tler Magazine, Inc., 565 F. Supp. 802, 803 (S.D. Tex. 1983) (magazine story
describing autoerotic asphyxiation). These cases should be distinguished from
the lawsuits over information that is harmful to third parties when acted on.
See, e.g., Rice v. Paladin Enter., Inc., 128 F.3d 233, 239 (4th Cir. 1997) (suit by
family of murder victims against publisher of HIT MAN: A TECHNICAL MANUAL
FOR INDEPENDENT CONTRACTORS). See generally Eugene Volokh, Crime-
Facilitating Speech, 57 STAN. L. REV. 1095, 1128–32 (2005) (discussing these
latter cases in great detail). The former involve dangerous falsehoods; the lat-
ter involve dangerous truths. See Susan M. Gilles, “Poisonous” Publications
and Other False Speech Physical Harm Cases, 37 WAKE FOREST L. REV. 1073,
1077–78 (2002) (making distinction). See generally Juliet Dee, “How-To” Man-
uals for Hitmen: Paladin Press, a Triple Murder, and First Amendment Protec-
tion of Technical Information, 23 COMM. & L. 1 (2001) (surveying caselaw on
dangerous information).
268. This distinction depends on the nature of their speech, not on the in-
stitutional status of Putnam or Jeppesen as the “press.” We are long past the
922 MINNESOTA LAW REVIEW [98:868
falsity has already been crossed—the only remaining argument
is over what degree of fault the First Amendment requires. The
same is true for search bias: false rankings made with actual
269
malice may be actionable.
C. SEARCH RANKINGS ARE FALSE WHEN THEY ARE DISHONEST
Now we are in position to apply law to fact. The problem of
falsity, it should be apparent, is a problem of baselines. To say
that a ranking is “false,” we need some baseline of truth. The
conduit, editor, and advisor theories diverge in the baselines
era in which it made sense—if it ever did—to think of the institutional media
as having a distinct and more protected First Amendment role. The Internet
makes it possible for anyone to publish, cheaply and quickly. This Article,
therefore, does not consider arguments that search engines are or are not
“press” or “news media” or “publishers.” Instead, it looks to generally applica-
ble First Amendment doctrine.
269. Another case cited by Volokh and Falk is Blatty v. New York Times
Co., 728 P.2d 1177 (Cal. 1986), in which the California Supreme Court held
that an author could not sue the New York Times for allegedly leaving his
book off its best-seller list. The court held that the list was not “of and concern-
ing” the plaintiff because it did not “refer to Blatty or his novel” expressly or
by implication. Id. at 1185. Doctrinally, this is bizarre: “of and concerning” is
not an element of tortious interference. See RESTATEMENT (SECOND) OF TORTS
§ 766B (1979) (giving action for intentional improper interference); id. § 767
(listing factors to determine whether interference is improper). The court im-
ported the of-and-concerning element from the defamation tort, believing that
Rosenblatt v. Baer, 383 U.S. 75 (1966), had constitutionalized the element.
Blatty, 728 P.2d at 1182. But this misreads Rosenblatt, which was a defama-
tion case. Rosenblatt, 383 U.S. at 77. Improper interference is its own well-
established tort, with its own doctrinal structure and its own safeguards.
Whether they are constitutionally sufficient must be assessed on their own
merits, rather than by converting improper interference into defamation by
another name.
Falsity and actual malice have been broadly constitutionalized, see, e.g.,
Hustler Magazine, Inc. v. Falwell, 485 U.S. 46, 56 (1988) (intentional infliction
of emotional distress); Bose Corp. v. Consumers Union, Inc., 466 U.S. 485, 488
(1984) (product disparagement), but not of-and-concerning, as the partial dis-
sent in Blatty noted. See Blatty, 728 P.2d at 1187–88 (Grodin, J., concurring
and dissenting). The majority asserted that of-and-concerning applies “to all
claims whose gravamen is the alleged injurious falsehood of a statement.” Id.
at 1182. But this too is wrong. Consider another cause of action premised on
injurious falsehood: false advertising. Section 43(a) of the Lanham Act makes
liable “[a]ny person who . . . in commercial advertising or promotion, misrepre-
sents the nature, characteristics, qualities, or geographic origin of his or her or
another person’s goods, services, or commercial activities.” 15 U.S.C.
§ 1125(a)(1) (2012) (emphasis added). Thus, a competitor may sue for false
statements about the plaintiff’s own products, even when the competitor is
nowhere mentioned. See, e.g., TrafficSchool.com, Inc. v. Edriver Inc., 653 F.3d
820, 826–28 (9th Cir. 2011) (competing driving school had Lanham Act stand-
ing to sue for the false suggestion that plaintiff’s website at dmv.org was affil-
iated with state agencies).
2014] SPEECH ENGINES 923
they choose. The conduit theory assumes that there is an objec-
tive baseline of relevance: it treats relevance as an objectively
observable property of a website. The editor theory, by contrast,
denies the possibility of any baseline at all: it treats relevance
as a wholly subjective expression on the search engine’s part.
The advisor theory, which recognizes that search rankings have
both objective and subjective properties, charts a middle course.
When a search engine is subjectively dishonest with its users—
when it returns results other than the ones it believes users
will find the most relevant—the search engine’s own secret,
undisclosed belief about relevance serves as the baseline
against which the falsity of the rankings it actually returns can
be measured. This Section takes up the three theories in turn.
1. The Conduit Theory: Objective Falsity
Start, once again, with the conduit theory. Google’s critics
regularly assert that rankings are falsifiable because relevance
is objective. Foundem, a price comparison website, claims, “[f]or
the query ‘compare prices shoei xr-1000’, Foundem is one of on-
270
ly two or three truly relevant pages,” and “[i]t clearly makes
no sense to exclude price comparison sites from these results
271
[for searches like [best price canon eos 500d]].” These
are claims that relevance for these queries is objectively deter-
minable, and that Google’s results are demonstrably wrong be-
cause they exclude Foundem. Kurt Wimmer argues that it
could be misrepresentation for Google to claim “that its own
272
services are the most relevant.” Nextag CEO Jeffrey Katz
claims, “In addition, Google often uses its prime real estate to
promote its own (often less relevant and inferior) products and
services, prohibiting companies from buying its best advertise-
273
ments.”
All of these arguments only make sense if relevance has an
external reality, but it does not. The critics are probably right
that the intentions behind some queries are reasonably unam-
270. Foundem’s Google Story, supra note 129.
271. Google Penalty Myths, SEARCHNEUTRALITY.ORG (Nov. 19, 2009),
http://www.searchneutrality.org/eu-launches-formal-investigation/myths
-surrounding-google-penalties.
272. Kurt Wimmer, The Proper Standard for Constitutional Protection of
Internet Search Practices 6 (unpublished white paper), available at http://
goo.gl/GkL0Y.
273. Katz, supra note 77. Nextag is a product search engine that competes
with Google’s own product search service.
924 MINNESOTA LAW REVIEW [98:868
274
biguous. But this does not imply that the best results for
those queries are similarly unambiguous. Some users will find
Foundem easier to use and more helpful than Nextag; others
will have the opposite reaction. Which of these product search
sites should come up first in searches for [canon eos 500d
compare prices]—or whether Google Shopping should—is
not a question with a unique answer. The diversity of users’
preferences for most queries will tend to make the choice to
rank one website over another nonfalsifiable.
2. The Editor Theory: No Falsity
Now for the editor theory. The Search King court wrote
that rankings are subjective “because every algorithm em-
ployed by every search engine is different, and will produce a
different representation of the relative significance of a particu-
275
lar web site.” Replace “relative significance of a particular
web site” with “the number of jellybeans it would take to fill
Soldier Field” and the fallacy is apparent. Search King could
come to court with better math, and Google’s “representation”
would be demonstrably false. Search King conflated users’
normative opinions about websites with search engines’ de-
scriptive opinions about which websites users will find rele-
vant.
The difference matters because the two kinds of opinions
are protected speech for different reasons and to very different
extents. Normative opinions are protected speech because we
have decided as a society to treat matters of taste and value as
questions of individual conscience rather than objective agree-
ment. Elizabeth Hand says that Rick Moody is “one of our best
276
writers,” while Dale Peck says that he is “the worst writer of
274. See JORIS VAN HOBOKEN, SEARCH ENGINE FREEDOM: ON THE IMPLI-
CATIONS OF THE RIGHT TO FREEDOM OF EXPRESSION FOR THE LEGAL GOVERN-
ANCE OF WEB SEARCH ENGINES 43–47 (P. Bernt Hugenholtz ed., 2012) (sug-
gesting that the possibility of “right” answers varies between navigational,
transactional, and informational queries); Search Quality Rating Guidelines
1.0, supra note 228, § 2.4, at 9 (same).
275. Search King v. Google Tech., Inc., No. CIV-02-1457-M, 2003 WL
21464568, at *3 (W.D. Okla. May 27, 2003).
276. Elizabeth Hand, Searching for Lost Time, WASH. POST, July 1, 2007,
http://www.washingtonpost.com/wp-dyn/content/article/2007/06/28/
AR2007062802184.html (reviewing RICK MOODY, RIGHT LIVELIHOODS: THREE
NOVELLAS (2007)).
2014] SPEECH ENGINES 925
277
his generation.” Allowing them both to have their say re-
278
spects personal autonomy while promoting social pluralism.
In contrast, freedom of expression for descriptive opinions
is an instrumental goal: it helps encourage the creation of bet-
ter and more accurate knowledge about the world. Whether
they are predictions about the future, like a weather report, or
claims about presently existing but uncertain facts, like a
hydrogeologist’s estimate of the reserves in an oil field, descrip-
tive opinions have in common that they are attempts to de-
scribe accurately the objective world as it is. As such, they can
be wrong: they are capable of being disproven by the right evi-
279
dence, at least in principle. That is, it is not the case that
search rankings are absolutely protected speech because they
express normative opinions. Instead, they enjoy only the weak-
er, more contingent protections afforded to descriptive opinions.
A search engine does not itself have a normative opinion about
which webpages are best; rankings do not express a search en-
280
gine’s own values. Google is not a book critic; when it links to
Dale Peck’s hatchet job of a review, Google has no particular
view about the truth of the matter asserted. It asserts only that
users will find Peck’s review relevant. In practice, it will gener-
ally be impossible for a court to conclude that Google’s asser-
281
tions of relevance are wrong. But that is because of the diver-
sity of users’ (normative) opinions and the difficulty of
measuring them, rather than because of the expressivity of
282
Google’s (descriptive) opinions.
277. Peck, supra note 237.
278. See Post, supra note 239, at 659 (“[A]ny government effort to penalize
false judgments in public discourse would in effect use the force of the state to
impose the standards of a specific community.”).
279. See Jefferson Cty. Sch. Dist. No. R-1 v. Moody’s Investor’s Servs., 175
F.3d 848, 853 (10th Cir. 1999) (endorsing, and garbling, the distinction);
Kathyrn Dix Sowle, A Matter of Opinion: Milkovich Four Years Later, 3 WM. &
MARY BILL RTS. J. 467, 577–79 (1994) (endorsing this distinction).
280. Some search engines appear to interject their own tastes and values,
such as SeekFind, whose mission is “to provide God-honoring, biblically based,
and theologically sound Christian search engine results in a highly accurate
and well-organized format.” SEEKFIND, http://www.seekfind.org (last visited
Oct. 12, 2013). But SeekFind is really just a vertical search engine in disguise.
SeekFind is not trying to impose its own Christian standard of relevance on
non-Christians; it is trying to satisfy Christian users’ explicitly Christian
standards of relevance.
281. See supra Part IV.C.1.
282. There is a helpful analogy to false advertising law. Any individual
shopper’s preference for brand A over brand B is unverifiable. But a survey
showing that four out of five shoppers preferred brand A to brand B is verifia-
926 MINNESOTA LAW REVIEW [98:868
3. The Advisor Theory: Subjective Falsity
The advisor theory leads us in a slightly different direction.
Since we are concerned about loyalty, we should worry most
about cases in which the search engine deliberately falls short
of the best it is capable of for users. That is, we are looking for
search rankings that depart from the search engine’s own as-
sessment of relevance. Milkovich explains the crucial doctrinal
distinction in tort:
For instance, the statement, “I think Jones lied,” may be provable as
false on two levels. First, that the speaker really did not think Jones
had lied but said it anyway, and second that Jones really had not lied.
It is, of course, the second level of falsity which would ordinarily serve
as the basis for a defamation action, though falsity at the first level
283
may serve to establish malice where that is required for recovery.
In a search ranking, the second-level statement (“this website
is not relevant”) is unprovable and unfalsifiable. But the im-
plicit first-level statement (“[Google believes] this website is not
relevant”) is false where Google believes otherwise. A classic
English fraud case, Edgington v. Fitzmaurice, put the point
quite nicely:
There must be a misstatement of an existing fact, but the state of a
man’s mind is as much a fact as the state of his digestion. It is true
that it is very difficult to prove what the state of a man’s mind at a
particular time is, but if it can be ascertained it is as much a fact as
anything else. A misrepresentation as to the state of a man’s mind is,
284
therefore, a misstatement of fact.
One might ask why subjective falsity should ever matter.
In many cases, it does not—it is not material to listeners.
Whether a normative opinion is subjectively true or false is ir-
relevant as a matter of law. Even though “I hate broccoli”
might be a lie, to let that lie form the basis of liability would
undo the right to dislike broccoli. Subjective falsity is also irrel-
ble. Only the latter will serve as sufficient basis to establish a comparative
claim; only the latter is subject to judicial second-guessing. The preferences
themselves are still subjective, but it is possible to make objective statements
that the consumers, in aggregate, have those preferences. See, e.g., Johnson &
Johnson Vision Care, Inc. v. 1-800 Contacts, Inc., 299 F.3d 1242 (11th Cir.
2002) (considering objective survey evidence about subjective consumer pref-
erences). The problem in search is that “Crest.com is relevant” is a far vaguer
statement than “consumers prefer Crest.”
283. Milkovich v. Lorain Journal Co., 497 U.S. 1, 20 n.7 (1990); see also
Virginia Bankshares, Inc. v. Sandberg, 501 U.S. 1083, 1092 (1991) (“Such
statements [of reasons, opinions, or beliefs] are factual in two senses: as
statements that the directors do act for the reasons given or hold the belief
stated and as statements about the subject matter of the reason or belief ex-
pressed.”).
284. Edgington v. Fitzmaurice, (1885) 29 Ch.D. 459 at 483 (Eng.).
2014] SPEECH ENGINES 927
evant where a descriptive opinion is objectively true. For exam-
ple, it does not typically suffice to make a statement actionable
285
under the securities laws. Subjective falsity standing alone is
immaterial to investors who care about the objective reality the
286
statement describes.
There is, however, an important category of descriptive
opinions for which subjective falsity is highly important, be-
cause it can be bootstrapped into a form of objective falsity.
Where a speaker holds herself out as having special expertise
in the subject matter, she represents that her descriptive opin-
287
ions are the product of superior skill and judgment. Where
her relationship to listeners is sufficiently close to invite reli-
ance, the result is not unlike an estoppel. Her own secret un-
disclosed beliefs provide a baseline against which the truth of
her statements can be measured.
There is a particularly illuminating analogy between
search rankings and credit ratings: both are numerical state-
ments about quality that combine a huge amount of knowledge
288
about the world into a single, ambiguous statement. Because
a rating is only a prediction about the probability of default,
and because that probability is expressed on a non-numerical
scale, it is difficult or impossible to prove a credit rating objec-
289
tively false. But this fact has not deterred courts from allow-
285. See Virginia Bankshares, 501 U.S. at 1096; Couture, supra note 236
(endorsing this rule).
286. See Couture, supra note 236 (criticizing courts’ “unsound materiality
analyses” in falsity cases).
287. See, e.g., Bily v. Arthur Young & Co., 834 P.2d 745, 768 (Cal. 1992)
(“Moreover, when a party possesses or holds itself out as possessing superior
knowledge or special information or expertise regarding the subject matter
and a plaintiff is so situated that it may reasonably rely on such supposed
knowledge, information, or expertise, the defendant's representation may be
treated as one of material fact.”).
288. The missing link between Search King and Milkovich is actually a
credit rating case, Jefferson Cnty. Sch. Dist. No. R-1 v. Moody’s Investor’s
Servs., Inc., 175 F.3d 848 (10th Cir. 1999). Search King cites Milkovich only by
way of embedded quotations from Jefferson County. Others who have noted
the connection between search engines and credit-rating agencies include
Mark Patterson, Manipulation of Product Ratings: Credit-Rating Agencies,
Google, and Antitrust, CPI ANTITRUST CHRON., Apr. 2012, and Andrew Car-
roll, Don’t Be Evil . . . Unless It Increases Revenue: What the Operation of Cred-
it Rating Agencies Can Teach Us About Google, 31 TEMP. J. SCI. TECH. &
ENVTL. L. 93 (2012).
289. See, e.g., Compuware Corp. v. Moody’s Investors Servs., Inc., 499 F.3d
520, 529 (6th Cir. 2007) (“A Moody’s credit rating is a predictive opinion, de-
pendent on a subjective and discretionary weighing of complex factors. We find
no basis upon which we could conclude that the credit rating itself communi-
928 MINNESOTA LAW REVIEW [98:868
290
ing some suits against credit rating agencies. The key is bad
291
faith. An actionable rating is not merely a bad prediction in
light of how things turned out or in light of the evidence availa-
ble, but one affirmatively given in knowing or reckless violation
292
of the rater’s own standards. These ratings are “false” be-
cause the rating agency has promised that they honestly repre-
293
sent its estimate of creditworthiness—but they do not. As one
court explained,
When a rating agency issues a rating, it is not merely a statement of
that agency’s unsupported belief, but rather a statement that the rat-
ing agency has analyzed data, conducted an assessment, and reached
a fact-based conclusion as to creditworthiness. If a rating agency
knowingly issues a rating that is either unsupported by reasoned
analysis or without a factual foundation, it is stating a fact-based
294
opinion that it does not believe to be true.
Credit ratings exist at all and are relied on by investors pre-
cisely because they express the best available proxy for an un-
observable aspect of reality: the rating agency’s judgment.
Thus, in credit-ratings cases, subjective honesty is all. Norma-
tively, if a credit rating is subjectively false, it is worthless, and
the arguments that it is worth protecting or plays a valuable
role in the financial markets vanish. And doctrinally, subjective
cates any provably false factual connotation.”).
290. See, e.g., Anschutz Corp. v. Merrill Lynch & Co., 785 F. Supp. 2d 799
(N.D. Cal. 2011); Abu Dhabi Com. Bank v. Morgan Stanley & Co., 651 F.
Supp. 2d 155, 175–76 (S.D.N.Y. 2009).
291. See, e.g., King Cnty. v. IKB Deutsche Industriebank AG, 751 F. Supp.
2d 652, 664–65 (S.D.N.Y. 2010) (“I have already ruled that plaintiffs stated a
claim for fraud against Fitch, which means plaintiffs have adequately pled
that (1) Fitch did not ‘genuinely and reasonably believe’ the ratings it issued
or that (2) those ratings were ‘without basis in fact’—i.e., that they did not
‘hold the opinions expressed by the ratings.’” (internal quotations omitted)).
292. Compare Anschutz, 785 F. Supp. 2d at 824 (“TAC may bring negligent
misrepresentation claims against the Rating Agencies if plaintiff alleges that
the Agencies did not honestly entertain the opinions about the ratings at the
time they were issued.”), with Plumbers Union Local No. 12 Pension Fund v.
Nomura Asset Acceptance Corp., 632 F.3d 762, 775 (1st Cir. 2011) (dismissing
misrepresentation claim because “tellingly, the complaint stops short of alleg-
ing expressly that the leadership of S & P or Moody's believed that their com-
panies’ ratings were false or were unsupported by models that generally cap-
tured the quality of the securities being rated”).
293. See In re Credit Suisse First Boston Corp., 431 F.3d 36, 48 (1st Cir.
2005) (“In cases premised on misstatements of opinion, however, the falsity
element, at a minimum, entails an inquiry into whether the statement was
subjectively false . . . .”).
294. Abu Dhabi Com. Bank v. Morgan Stanley & Co., 888 F. Supp. 2d 431,
455 (S.D.N.Y. 2012).
2014] SPEECH ENGINES 929
falsity marks the dividing line that suffices to make a bad-faith
295
rating actionable.
And so back to search results. Google holds out to the
296
world that its rankings attempt to maximize relevance. In-
deed, the head of Google’s search ranking team wrote, “[o]ur al-
gorithms rank results based only on what the most relevant
297
answers are for users.” In pursuing this broad goal, Google,
like the ratings agencies, is free to establish its own criteria for
298
measuring and describing quality. It is not free, however, to
assert that it has attempted to maximize quality when it has
295. See Letter from Laurence H. Tribe & Thomas C. Goldstein to Sec. and
Exch. Comm’n 4 (Dec. 14, 2009), available at http://www.sec.gov/comments/s7
-24-09/s72409-13.pdf (“But [nationally recognized statistical rating organiza-
tions (NRSROs)] are not altogether immune from suit, for the First Amend-
ment does not preclude the imposition of liability for a factual misstatement in
a rating publication (including the fact that the NRSRO believes a certain rat-
ing opinion) if it is made with actual malice . . . .” (emphasis added) (internal
quotation marks omitted)); cf. Letter from Eugene Volokh to the Subcomm. on
Capital Mkts., Ins., and Gov’t Sponsored Enter. of the House Comm. on Fin.
Servs. (May 15, 2009), available at http://financialservices.house.gov/media/
file/hearings/111/volokh.pdf (calling credit ratings “pure opinion”). For further
academic analyses applying the First Amendment to credit ratings, see, for
example, Caleb Deats, Talk That Isn’t Cheap: Does the First Amendment Pro-
tect Credit Rating Agencies’ Faulty Methodologies from Regulation?, 110
COLUM. L. REV. 1818 (2010); Parisa Haghshenas, Obstacles to Credit Rating
Agencies’ First Amendment Defense in Light of Abu Dhabi, 8 FIRST AMEND. L.
REV. 452 (2010); Jonathan W. Heggen, Not Always the World’s Shortest Edito-
rial: Why Credit-Rating-Agency Speech Is Sometimes Professional Speech, 96
IOWA L. REV. 1745 (2011); Gregory Husisian, What Standard of Care Should
Govern the World’s Shortest Editorials?: An Analysis of Bond Rating Agency
Liability, 75 CORNELL L. REV. 410 (1990); Kenneth C. Kettering, Securitization
and Its Discontents: The Dynamics of Financial Product Development, 29
CARDOZO L. REV. 1553, 1689–91 (2008); Theresa Nagy, Note, Credit Rating
Agencies and the First Amendment: Applying Constitutional Journalistic Pro-
tections to Subprime Mortgage Litigation, 94 MINN. L. REV. 140 (2009); Ulrich
G. Schroeter, Three Letters That Move the Markets: Credit Ratings Between
Market Information and Legal Regulation, J. APPLIED RES. ACCT. & FIN., July
2011, at 14.
296. See Schmidt Testimony, supra note 16, at 14 (“At all times, Google’s
primary motivation has been improving the search experience for our users by
providing the most relevant and useful information in response to their que-
ries.”).
297. Singhal, supra note 223.
298. In particular, it will usually be free to make either choice when it has
to choose between results that will satisfy some users and results that will sat-
isfy others. Even though Google knows that some users prefer Nextag and
others prefer Foundem, distinguishing them and giving them each what they
want will be difficult or impossible. Subjective dishonesty arises only when the
search engine, by its own lights, concludes that it really “ought” to return
Nextag and returns Foundem instead.
930 MINNESOTA LAW REVIEW [98:868
299
not actually done so. That is a false statement of fact, one
implicitly embedded in every ranking it utters that is based on
something other than relevance. Determining whether Google
believes its search rankings, of course, requires looking at its
surveys of user relevance assessments, its internal treatment of
those surveys, and its procedures for translating those assess-
300
ments into search rankings.
What is more, the bad faith in misrepresenting the process
by which search results are generated can also suffice to
demonstrate the necessary fault. Consider again the quoted
passage from Milkovich. If Google does not believe the rankings
it provides to a user, this is “falsity at the first level”—the
equivalent of a “speaker [who] really did not think Jones had
301
lied but said it anyway.” But as the Court noted, this dishon-
302
esty “may serve to establish malice.” It is precisely because
the speaker has direct access to her own beliefs that a false
statement about them is knowingly false. So too with a search
engine. If Google deliberately changes its rankings to make
them less relevant, the results are not just false, but knowingly
false. If Google changes its rankings with deliberate indiffer-
ence to relevance and the results are in fact less relevant by its
usual standards, they are not just false, but recklessly false. Ei-
ther way, subjective bad faith in reporting rankings constitutes
actual malice. This is an important convergence. The combina-
tion of subjectivity and objectivity in a search ranking mean
299. In theory, the falsehood could be either general (because the ranking
criteria Google uses are not an honest attempt to implement relevance), or
specific (because a given ranking did not actually result from the application of
those criteria). In practice, however, the distinction collapses, as described in
Part IV.D infra. Since no significant consequences turn on it, it is better not to
attempt the difficult exercise in line-drawing involved.
300. If this sounds like a difficult exercise, consider that it may still be an
easier hill to scale than objective falsity. See generally Rebecca Tushnet, It De-
pends on What the Meaning of “False” Is: Falsity and Misleadingness in Com-
mercial Speech Doctrine, 41 LOY. L.A. L. REV. 227 (2007) (discussing the diffi-
culty in making judgments about the objective falsity of advertisements). It
may also be difficult for private plaintiffs to frame a complaint that complies
with Rule 11 and alleges sufficient facts to provide a plausible inference of
subjective dishonesty: the relevant facts about a search engine’s relevance as-
sessments and algorithms will typically be closely guarded secrets within the
search engine’s possession. FTC oversight, including FTC subpoena power,
may be necessary if one is committed to stamping out subjective falsity. See
James Grimmelmann, Devils and Details, THE LABORATORIUM (Jan. 4, 2013,
1:21 AM), http://laboratorium.net/archive/2013/01/04/devils_and_details. Some
readers may not consider the game to be worth the candle.
301. Milkovich v. Lorain Journal Co., 497 U.S. 1, 20 n.7 (1990).
302. Id.
2014] SPEECH ENGINES 931
that falsity and fault are not just connected, but coextensive.
The same facts that establish one establish the other.
This analysis does not depend on the fact that Google ex-
plicitly embraces relevance as its goal. The social practice of
search is oriented around relevance: even in the absence of ex-
plicit claims, users would reasonably assume that Google is not
deliberately hiding relevant results. Defamation law takes the
common-sense position that speakers ordinarily have bases for
their statements, so a statement of “opinion” with no stated ba-
sis may sometimes be treated as implicitly asserting the exist-
ence of some underlying facts sufficient to warrant the opin-
303
ion. So too with search. Users would naturally expect that the
search engine has a relevance-related justification for returning
the results it does and not others.
How far can explicit disclaimers go? Could Google draft a
disclaimer that would entirely exonerate it from deception-
based claims like tortious interference? Search without rele-
vance is pointless, so a complete disclaimer of relevance would
be self-evidently false. Rather, to be effective, a disclaimer
would need to affirmatively reveal the other considerations en-
tering into a ranking, such as legal compliance, protection of
good morals, or the desire to crush Larry Page’s enemies, see
them driven before him, and hear the lamentations of their
women.
* * *
To recap, the conduit theory goes wrong because it treats
relevance as a fact about websites and ignores users’ normative
opinions of websites. The editor theory goes wrong because it
conflates users’ normative opinions about websites with search
engines’ descriptive opinions about which websites users will
find relevant. In contrast, the advisor theory of relevance—a
descriptive opinion about users’ normative opinions of web-
site—yields a straightforward test based on loyalty to the user.
A search ranking is actionable in tort when it is subjective-
304
ly dishonest. A ranking is meaningfully false when it is given
303. See, e.g., TMJ Implants, Inc. v. Aetna, Inc., 498 F.3d 1175, 1183–87
(10th Cir. 2007) (discussing the Restatement test in detail and concluding,
“[i]n sum, we find little difference between § 566 and the Milkovich standard”);
RESTATEMENT (SECOND) OF TORTS § 566 (1977).
304. The following is a sketch of how the analysis of regulations designed
to protect search users’ interests as active listeners may be slightly different.
It is offered tentatively, as a starting point for discussion.
In Search King-style suits, the basis for liability is deception. But just as
consumer-protection and false advertising law can go beyond deceptive state-
932 MINNESOTA LAW REVIEW [98:868
in knowing or reckless disregard of the search engine’s own in-
ternal standards for evaluating users’ relevance judgments.
Such bad-faith rankings will also automatically satisfy
Milkovich’s actual-malice standard of fault. Falsity and fault
converge for search results. This is an attractive compromise be-
tween the editor theory, which asserts that rankings can never
be false, and the conduit theory, which treats falsity as trivial
when it bothers to worry about falsity at all.
D. ALGORITHMS ARE A RED HERRING
Some commentators would say that the entire above anal-
ysis is misguided because no human communication is involved
at all. They believe it makes a significant difference that Google
uses computers to generate its search results. They are
305
wrong.
Tim Wu argues in yet another New York Times op-ed about
Google that “computerized decisions” should not be considered
306
speech. Comparing search results to GPS route suggestions,
Microsoft Word spell-checking, and Facebook friend sugges-
tions, he explains that “computer programs are utilitarian in-
struments meant to serve us” whereas the First Amendment “is
intended to protect actual humans against the evil of state cen-
307
sorship.” He concludes:
The line can be easily drawn: as a general rule, nonhuman or au-
tomated choices should not be granted the full protection of the First
Amendment, and often should not be considered “speech” at all.
ments to misleading ones, even where the statements are purportedly ex-
plained away in the fine print, so too search regulation could reasonably pro-
hibit practices that reduce relevance, even where they are purportedly dis-
closed in the fine print. As with commercial speech, search rankings are
covered speech because of their value to listeners, and regulations to protect
those listeners are consistent with the level of protection they enjoy. If this is
right, then the government could make a prima facie case for a regulation by
showing that the practice it prevents is not directed to improving relevance for
users, shifting an evidentiary burden to search engines to show that it has
(possibly longer-term) sufficient relevance-improving effects.
Thus, relevance could potentially provide a baseline against which ma-
nipulation, and not just deception, could be defined. But this would require
care—more care than most of Google’s critics have taken—and the doctrinal
details would be subtle. Given the closure of the FTC’s investigation, it is a
subject for another day.
305. See generally Benjamin, supra note 220 (arguing that many “algo-
rithm-based outputs” are speech for First Amendment purposes).
306. Tim Wu, Op-Ed., Free Speech for Computers?, N.Y. TIMES, June 20,
2012, at A29.
307. Id.
2014] SPEECH ENGINES 933
(Where a human does make a specific choice about specific content,
308
the question is different.)
Wu’s argument misses the nature of the (very human)
opinions expressed in search results, because it slights the idea
309
that opinions can be expressed through automated processes.
If Google consisted of Larry Page sitting at a computer person-
ally typing out answers to users’ queries, his responses would
constitute protected speech. The actual Google differs because
Page and his employees have written a complicated computer
program that takes users’ queries as its input and produces
search results as its output.
This is not a meaningful distinction when thinking about
search results. Suppose that Larry Page programs a computer
to respond “Try Bobo’s Drive-In” whenever a user types in
[food]. In Wu’s terms, this is a “specific choice about specific
content” made repeatedly; it walks and quacks like speech. The
same reasoning applies to any other up-front programming
choice that has numerous predictable consequences: directing
users to restaurants’ own webpages over review sites (or vice-
versa), directing GPS users to take arterial roads rather than
side streets (or vice-versa), or directing the friend suggester not
to suggest as friends people currently in a relationship with us-
ers’ exes. Whether or not each of these decisions is speech, algo-
rithmically multiplying its consequences a millionfold should
not change the answer.
Nor can “the algorithm” be used as a baseline from which
310
any deviation is impermissible deception. There is no mean-
308. Id.
309. For versions of this reply to Wu, see Timothy B. Lee, Do You Lose Free
Speech Rights If You Speak Using a Computer?, ARS TECHNICA (June 22,
2012, 11:15 AM), http://arstechnica.com/tech-policy/2012/06/do-you-lose-free
-speech-rights-if-you-speak-using-a-computer/; Eugene Volokh, Freedom of
Speech and Information Produced Using Computer Algorithms, VOLOKH CON-
SPIRACY (June 21, 2012, 11:59 PM), http://www.volokh.com/2012/06/21/
freedom-of-speech-and-information-produced-using-computer-algorithms/. For
a more comprehensive treatment of the algorithmic speech question, see Ben-
jamin, supra note 220; see also Bruce Boyden, Speech by Proxy, MADISONIAN
(June 25, 2012), http://madisonian.net/2012/06/25/speech-by-proxy. Even Tim
Wu has since endorsed a narrower and more careful theory of algorithmic
speech. See Tim Wu, Machine Speech, 161 U. PA. L. REV. 1495, 1530–31 (2013)
(recognizing that “Google is just trying to find what the user wants” and en-
dorsing a First Amendment test that “[t]he law in question must somehow
burden a search engine’s users”). This is not quite complete—it leaves out the
way in which Google communicates valuable opinions about relevance to us-
ers—but it recognizes the crucial role played by search users.
310. See, e.g., Grimmelmann, supra note 1, at 59–60 (describing “strong
934 MINNESOTA LAW REVIEW [98:868
ingful dividing line between algorithmic and manual search re-
sults. Every search result is produced algorithmically; it’s algo-
rithms all the way down. Compare two cases involving Google’s
responses to rankings manipulation. In November 2010, the
New York Times reported on DecorMyEyes, an online glasses
vendor that cheated its customers and then deliberately offend-
ed them when they complained, knowing that they would post
furious reviews—thereby tricking Google into thinking that
311
DecorMyEyes was a popular site. In February 2011, the same
312
reporter caught a much bigger fish: J.C. Penney. The retailer
had engaged in “the most ambitious attempt” at gaming Google
an industry expert had ever seen, buying thousands of links to
JCPenney.com from unrelated websites. Google “developed an
algorithmic solution” to detect and demote hundreds of mer-
chants like DecorMyEyes that “provide an extremely poor user
313
experience.” In contrast, Google took individual “manual ac-
tion” against J.C. Penney, dropping its website from the num-
ber-one result for “living room furniture” to number sixty-
314
eight. There is no meaningful difference between the cases:
relevance is the real issue, not algorithmic versus manual
ranking.
E. CONCLUSION: THE FEDERAL TRADE COMMISSION GETS IT
MOSTLY RIGHT
When the FTC closed its search-bias investigation into
Google, it seems to have acted consistently with its mission as a
consumer-protection agency and recognized that Google’s users
were the real parties in (the public) interest. The FTC’s official
statement explained, in essence, that Google’s results are not
subjectively false because Google’s algorithms are a good-faith
effort to maximize user relevance:
The totality of the evidence indicates that, in the main, Google adopt-
ed the design changes that the Commission investigated to improve
the quality of its search results, and that any negative impact on ac-
tual or potential competitors was incidental to that purpose. . . .
intuitive appeal” of manual manipulation argument).
311. David Segal, A Bully Finds a Pulpit on the Web, N.Y. TIMES, Nov. 28,
2010, at BU1.
312. David Segal, The Dirty Little Secrets of Search, N.Y. TIMES, Feb. 13,
2011, at BU1.
313. Amit Singhal, Being Bad to Your Customers Is Bad for Business,
GOOGLE OFFICIAL BLOG (Dec. 1, 2010), http://googleblog.blogspot.com/2010/12/
being-bad-to-your-customers-is-bad-for.html.
314. Segal, supra note 312.
2014] SPEECH ENGINES 935
While Google’s prominent display of its own vertical search results
on its search results page had the effect in some cases of pushing oth-
er results “below the fold,” the evidence suggests that Google’s prima-
ry goal in introducing this content was to quickly answer, and better
satisfy, its users’ search queries by providing directly relevant infor-
315
mation.
Google’s critics were outraged, but this was probably the right
result. The FTC asked the right question (“Did Google adjust
its algorithms for the purpose of sending users to less relevant
sites?”) and came to a defensible answer (“No.”).
The advisor theory’s focus on falsity, which necessarily re-
quires an evaluation of the search engine’s good faith, explains
the FTC’s otherwise surprising turn toward considering mo-
316
tives. Its analysis seems to slide back and forth between dis-
cussion of Google’s motives and the effects on consumers. But
the two are inextricably linked; Google acted in good faith be-
cause its own studies showed that the changes benefitted users.
“Reasonable minds may differ” about search results, the FTC
wrote, and its decision properly preserved a safe space for de-
317
ductive opinions about relevance.
Indeed, many of Google’s seemingly problematic practices
can be defended, sometimes quite convincingly, as good-faith
enhancements to relevance. The penalty that Google applied to
Foundem and other price-comparison sites reduces the promi-
nence of dozens of me-too sites with little to distinguish one
from another. The same goes for Google’s decisions to devote
front-page search space to Google+, Universal Search, and
Knowledge Graph results. Google can quite reasonably believe
that integrating its affiliated sites into results is relevance-
improving overall. One may disagree—the present author
thinks that the Google+ integration is bad for users and bad for
Google—without believing that these decisions, on the evidence
available, amount to bad faith.
315. Statement of the Fed. Trade Comm’n, supra note 5, at 2 (emphasis
added).
316. The FTC added that the changes probably benefitted consumers, but
it is striking that the Commission put the intent first and made the actual ef-
fects the corollary. To underscore the anomaly of this approach, consider that
some of Google’s strongest antitrust defenders are also fierce critics of the use
of intent evidence in antitrust cases. Compare Geoffrey A. Manne & Joshua D.
Wright, Google and the Limits of Antitrust: The Case Against the Case Against
Google, 34 HARV. J.L & PUB. POL’Y 171 (2011) (three cheers for Google), with
Geoffrey A. Manne & Joshua D. Wright, Innovation and the Limits of Anti-
trust, 6 J. COMPETITION L. & ECON. 153 (2010) (three jeers for intent evidence).
317. Statement of the Fed. Trade Comm’n, supra note 5, at 3.
936 MINNESOTA LAW REVIEW [98:868
If there is a fly in the ointment, it is that while the FTC
318
pledged to “remain vigilant and continue to monitor Google”
it did not give much thought as to how to carry out its monitor-
ing. Since search bias claims hinge on Google’s honesty in fol-
lowing its own processes, an outside observer will rarely have
the necessary information to reliably conclude that something
fishy is taking place. Only the FTC, with its subpoena power, is
319
well positioned to look “under the hood.” Dropping the com-
plaint entirely, as the FTC did, abdicates that responsibility.
Some kind of regular ongoing opening up of the algorithms is
the most effective way to keep Google loyal. Given Google’s size
and significance, the FTC should have given more thought to
setting a continuing compliance regime—like the ones that
320
credit rating agencies are required to have in place. This isn’t
just about Google: Bing is big enough, and potentially bad
enough, that it ought to have ongoing oversight too.
V. OTHER APPLICATIONS
The advisor theory is useful well beyond search bias.
Google’s critics allege that it infringes copyright on an epic
scale, tramples user privacy, smears the innocent, and kicks
puppies for fun. Google, needless to say, sees matters rather
differently. On the company’s account, it is guilty only of offer-
ing the best search results on the web, bar none. The advisor
theory helps sort problematic practices from benign ones. This
Part gives brief sketches of four other legal controversies
around search. It does not offer comprehensive analyses of any
of the controversies; that will need to wait for future work. In-
stead, it shows how access and loyalty—the key prescriptions of
the advisor theory—bring fresh insights to well-worn disputes.
318. Id. at 4.
319. Bracha & Pasquale, supra note 56, at 1202.
320. See 15 U.S.C. § 78o-7(c)(3) (2012) (requiring each NRSRO to “estab-
lish, maintain, enforce, and document an effective internal control structure
governing the implementation of and adherence to policies, procedures, and
methodologies for determining credit ratings”); 17 C.F.R. § 240.17g-2 (2013)
(imposing record-keeping requirements on registered NRSROs); id. § 240.17g-
6(a)(2)–(3) (prohibiting an NRSRO from “[i]ssuing . . . a rating that is not de-
termined in accordance with the nationally recognized statistical rating organ-
ization's established procedures and methodologies” in certain cases involving
a potential conflict of interest); see also Carroll, supra note 288, at 116–18 (en-
dorsing Dodd-Frank-ization of Google).
2014] SPEECH ENGINES 937
A. COPYRIGHT
Google dreams big, and none of its dreams are bigger than
321
its plan to scan every book ever published. Not the driverless
322 323
cars. Not the virtual-reality glasses. Not even the prize to
324
land a robot on the moon. No, Google Books—a program regu-
larly compared to a modern Library of Alexandria—best cap-
325
tures the company’s ambition and arrogance. Google borrows
physical books from libraries and digitizes them, then feeds the
texts into its search engine, which tells users who wrote that,
and on what page. So far, Google is up to about 20 million
326
books. The program has drawn four separate lawsuits by au-
327
thors and publishers for copyright infringement. These are
lawsuits over indexing: their goal is to stop Google from putting
content in its search index without the provider’s permission.
The conduit theory would say that since search is a vehicle
for websites and other publishers to be found, it follows that
they ought to be findable on their own terms. A search engine
should be required to index them when they want to be includ-
ed; and required not to index them when they want to be ex-
cluded. On this view, Google Books should have been confined
to voluntary agreements with authors and publishers. And the
editor theory is ambiguous. A newspaper exercises editorial
judgment in choosing comics and columnists—but it needs
permission from the authors of both to print them. On the other
hand, when the newspaper reports on goings on around town, it
is organizing and delivering content of its own, not simply re-
packaging the content of the art galleries and theaters. On this
321. See Jeffrey Toobin, Google’s Moon Shot, NEW YORKER, Feb. 5, 2007, at
30.
322. See John Markoff, Google Cars Drive Themselves, in Traffic, N.Y.
TIMES, Oct. 10, 2010, at A1.
323. See Nick Bilton, A Rose-Colored View May Come Standard, N.Y.
TIMES, Apr. 5, 2012, at B3.
324. See John Schwartz, $25 Million in Prizes Is Offered for Trip to Moon,
N.Y. TIMES, Sept. 14, 2007, at A14.
325. See Sergey Brin, Op-Ed., A Library to Last Forever, N.Y. TIMES, Oct.
9, 2009, at A31.
326. See Jennifer Howard, Google Begins to Scale Back Its Scanning of
Books from University Libraries, CHRON. HIGHER ED., Mar. 9, 2012, http://
chronicle.com/article/Google-Begins-to-Scale-Back/131109/.
327. See Authors Guild v. HathiTrust, 902 F. Supp. 2d 445 (S.D.N.Y. 2012);
Authors Guild v. Google, Inc., 282 F.R.D. 384 (S.D.N.Y. 2012); Authors Guild
v. Google, Inc., 770 F. Supp. 2d 666 (S.D.N.Y. 2011); Complaint, McGraw-Hill
Cos. et al. v. Google, No. 05-cv-08881, 2005 WL 2778878 (S.D.N.Y. Oct. 19,
2005).
938 MINNESOTA LAW REVIEW [98:868
latter view, Google Books is a wholly new product, one distinct
328
from the books it scans.
The advisor theory comes down decisively in favor of index-
ing. Users’ interests cut uniformly in favor of maximizing the
universe of searchable information. Indexing is purely an issue
of access: a search engine never acts disloyally by indexing
more content. A provider who insists on structuring how users
learn about its information is, in essence, taking control of the
search process through vertical integration. This limits users’
choices among search technology—and directly inhibits their
ability to compare among providers. There is rarely a good rea-
son for a speaker to be willing to share its speech with listeners
while preventing them from knowing about it. Taking the us-
er’s point of view emphasizes the enormous societal gains from
searchability: entirely new ways of finding and learning from
329
works become possible.
Thus, the advisor theory is even more radically pro-
indexing than the editor theory. For search engines, indexing is
a business decision; for users, it is an essential precondition to
informational freedom. Indeed, there is a strong argument that
information has not been meaningfully “published” until it is
made searchable. This is the position taken by patent law: a
thesis in a library does not qualify as prior art until it is not
330
just physically accessible to the public but properly indexed.
For copyright purposes, anything openly published should be
searchable. That requires a blanket privilege to copy for the
purposes of indexing, and a privilege to show excerpts to users
to help them decide whether to follow up on search results by
consulting the original.
One important exception may be privacy. Think of a father
who puts photographs of his daughter online and emails the
link to family members for sharing with their friends, but who
331
prefers not to have the pictures show up in search engines. A
328. See Matthew Sag, Copyright and Copy-Reliant Technology, 103 NW. U.
L. REV. 1607 (2009) (endorsing broad fair-use protections for such technolo-
gies).
329. See Matthew Sag, Orphan Works as Grist for the Data Mill, 27 BERK.
TECH. L.J. 1503 (2012).
330. Compare In re Cronyn, 890 F.2d 1158 (Fed. Cir. 1989) (thesis indexed
only by title on index card in shoebox is not a “printed publication”), with In re
Hall, 781 F.2d 897 (Fed. Cir. 1986) (thesis indexed in library catalog is a
“printed publication”).
331. See Lauren Gelman, Privacy, Free Speech, and “Blurry-Edged” Social
Networks, 50 B.C. L. REV. 1315 (2009); Woodrow Hartzog & Fred Stutzman,
2014] SPEECH ENGINES 939
privacy exception, however, makes less sense for books than it
does for webpages. Google offers authors and publishers an opt-
332
out from book scanning, just as it offers websites an opt-out
333
from its main search engine. It’s not clear that Google needs
to offer opt-outs for books, or that it should.
Two recent decisions show how fair use can be calibrated to
accommodate indexing. In Authors Guild, Inc. v. HathiTrust, it
was a fair use for Google’s partner libraries to use their copies
334
of the scanned books to create their own search engine. But
in Associated Press v. Meltwater Holdings, Inc., it was not a fair
use for a news monitoring service to send reports to its sub-
scribers containing substantial excerpts from news stories pub-
335
lished on the web. Both drew on a line of cases finding that
search engines make transformative fair uses of the material
they index because the search engine serves a different purpose
336
than the works it describes. The purpose is only different
from the user’s point of view: she consults the search engine to
find the works, and consults the works themselves to experi-
ence and understand them. The HathiTrust court embraced the
search-engine cases, saying it “cannot imagine a definition of
fair use that would not encompass the transformative uses
337
made by” the libraries. But the Meltwater court held that
Meltwater was not a search engine, because it was a “subscrip-
338
tion service” rather than a “publicly available tool” and be-
cause its searches were “run against a defined list of content
339
providers” rather than the Internet as a whole. Both distinc-
tions are singularly unpersuasive: the court’s distinctions
The Case for Online Obscurity, 101 CALIF. L. REV. 1, 35–37 (2013).
332. See Information for Publishers and Authors About the Library Project,
GOOGLE BOOKS, http://books.google.com/googlebooks/publisher_library.html
(last visited Nov. 27, 2013).
333. See Field v. Google, Inc., 412 F. Supp. 2d 1106, 1112–13 (D. Nev. 2006)
(stating that since Google cannot contact all website owners, they can “auto-
matically communicate their preferences” by putting “instructions in
‘metatags’”).
334. Authors Guild, Inc. v. HathiTrust, 104 U.S.P.Q.2d (BNA) 1659, 1669
(S.D.N.Y. 2012).
335. Associated Press v. Meltwater U.S. Holdings, Inc., 931 F. Supp. 2d 537
(S.D.N.Y. 2013).
336. See Perfect 10, Inc. v. Amazon.com, Inc., 508 F.3d 1146, 1164–65 (9th
Cir. 2007) (image thumbnails transformative); Kelly v. Arriba Soft Corp., 336
F.3d 811, 818 (9th Cir. 2003) (image thumbnails transformative); Field, 412 F.
Supp. 2d, at 1118–19 (search engine cache transformative).
337. HathiTrust, 104 U.S.P.Q.2d, at 1675.
338. Meltwater, 931 F. Supp. 2d at 554.
339. Id. at 555.
940 MINNESOTA LAW REVIEW [98:868
would require all search engines to support themselves with
advertising and would prohibit vertical search entirely.
The better distinction between the two cases has to do with
how users employed the two services. Obtaining a list of books
containing a search term is only the first step in the research
process; to learn more, one must still obtain a copy and read
the book. That’s precisely the kind of connection between au-
thors and readers that the copyright system encourages; it’s
beyond perverse for authors to object that a search engine rec-
ommends their works. But the Associated Press made a plausi-
ble argument that Meltwater’s users were using its clippings as
a substitute for reading the original stories; it had a click-
340
through rate of less than a tenth of a percent. The details are
debatable, but the general principle Meltwater embraces is
sound: search engines have a better fair use case when they
help users find websites than when they merely republish web-
341
sites’ content.
On November 14, 2013, as this Article was going through
the final stages of editing for publication, the Southern District
of New York held that it was also a fair use for Google to use its
own scans to offer the Google Books search engine and to show
“snippets” of one eighth of a page around each search result to
342
users. The court’s conclusion was consistent with the argu-
ment above: “snippets help users locate books and determine
whether they may be of interest. . . . Google Books does not su-
persede or supplant books because it is not a tool to be used to
343
read books.” The court concluded that Google Books “advanc-
es the progress of the arts and sciences, while maintaining re-
spectful consideration for the rights of authors and other crea-
tive individuals, and without adversely impacting the rights of
344
copyright holders. . . . Indeed, all society benefits.”
340. Id. at 547. There are multiple reasons why this could be true, not all of
them infringing. The court noted that Meltwater showed its users excerpts of
up to 440 characters, representing between 4.5% and 61% of the AP’s articles.
Id. at 545–47. But it did not persuasively explain whether the low click-
through-rate was attributable to Meltwater’s display of AP’s protectable ex-
pression, rather than because it was relaying the uncopyrightable underlying
facts.
341. See id. at 552.
342. See Authors Guild v. Google, Inc., No. 05 Civ. 8136 (DC) (S.D.N.Y.
Nov. 14, 2013).
343. Id. at *20–21.
344. Id. at *26.
2014] SPEECH ENGINES 941
B. PRIVACY
Search is valuable, but it is not free. To generate individu-
ally meaningful results, the search engine requires access to
the personal information that distinguishes one user from an-
other. The current query is just the tip of the iceberg: over time,
a search engine can accumulate an extensive profile of a user’s
interests. This intellectual history can be intensely personal
345
and immensely revealing. Search privacy is therefore a sub-
ject of significant concern for consumer advocates; some search
engines even compete by emphasizing that they retain less in-
346
formation on users.
Neither the conduit nor the editor theory is much help
here; their attention is elsewhere. Neither transmitting website
speech nor curating a collection of links has any necessary con-
nection to user information. Thus, both theories treat any flow
of information from the user to the search engine as a separate
issue from the quality of search results.
But on a user-centric view, user data takes center stage. It
is the search query that defines search: with no query, the
search engine has no question to answer. The very thing that
makes search sensitive to user interests means that search en-
gines also acquire sensitive information about what users are
347
interested in. There is no way to engineer a search engine
that does not observe user interests. And from a user’s perspec-
345. See, e.g., Michael Zimmer, Privacy on Planet Google: Using the Theory
of “Contextual Integrity” to Clarify the Privacy Threats of Google’s Quest for the
Perfect Search Engine, 3 J. BUS. & TECH. L. 109, 112–14 (2008). The record-
holder here may be user “927,” whose imperfectly anonymized queries were
released along with 650,000 others’ by AOL in a well-publicized 2006 scandal.
See Michael Barbaro & Tom Zeller, Jr., A Face Is Exposed for AOL Searcher
No. 4417749, N.Y. TIMES, Aug. 9, 2006, at A1. This user’s queries included
[cut into your trachea], [beauty and the beast beastility
porn], [holocaust rape], [was abe lincoln gay], [intersexed
genetails], and [low carb calorie foods]. Id. To read 927’s and other
users’ query histories is to wince at the consequences if they were to be linked
back to specific individuals. Which may be surprisingly easy. See generally
Paul Ohm, Broken Promises of Privacy: Responding to the Surprising Failure
of Anonymization, 57 UCLA L. REV. 1701 (2010) (explaining the ease of re-
identifying supposedly anonymous individuals).
346. DuckDuckGo, for example, has a charming explanation of its policies
against user tracking. DUCKDUCKGO, http://donttrack.us/ (last visited Nov.
27, 2013).
347. Cf. Nathan Newman, The Cost of Lost Privacy: Search, Antitrust and
the Economics of the Control of User Data (May 14, 2013) (unpublished manu-
script), available at http://papers.ssrn.com/sol3/papers.cfm?abstract_id=
2265026 (arguing that the exploitation of personal information is so central to
Google’s business practices that it amounts to an antitrust violation).
942 MINNESOTA LAW REVIEW [98:868
tive, it is also a matter of some importance what is done with
that information once it has been handed over to the search en-
gine. A search for [san jose jobs in sales] or [furry
videos] could be embarrassing or worse in the wrong hands.
348
Searches implicate intellectual privacy, which goes to the
heart of users’ ability to lead autonomous self-directed lives by
349
forming their own private opinions about the world. The
freedom to think for oneself requires the freedom to read unob-
350
served, which in turn requires the freedom to search unob-
served.
In agency terms, an agent has a duty not to misuse confi-
351
dential information supplied by the principal. This duty can
be waived with properly informed consent, but the common-law
baseline is that an agent or advisor in a fiduciary relationship
352
must respect client confidences. Thus, the debate over search
user privacy ought to start from this baseline: query data and
other data supplied by the user as part of obtaining search re-
sults are subject to a duty of confidentiality. Search engines
may not transfer any of this data to third parties without in-
formed consent. Nor may they use it against the interests of
their principals—search users—without informed consent.
This last point has important implications for the gold
mine at the heart of Google’s advertising business, which is
353
based on precisely targeted advertising. Some of this target-
ing is valuable to users and valued by them: showing geograph-
ically targeted florist ads on a search for [flowers] is another
way of improving relevance. But in its more comprehensive and
intrusive forms, targeted advertising raises serious autonomy
348. See Neil Richards, Intellectual Privacy, 87 TEX. L. REV. 387 (2008).
349. See M. Zimmer, The Gaze of the Perfect Search Engine: Google as an
Infrastructure of Dataveillance, in WEB SEARCH: MULTIDISCIPLINARY PER-
SPECTIVES 77, 77 (Amanda Spink & Michael Zimmer eds., 2008) (arguing that
search engine surveillance of users “threaten[s] the values the perfect search
engines were designed to sustain”).
350. See Julie E. Cohen, A Right to Read Anonymously: A Closer Look at
“Copyright Management” in Cyberspace, 28 CONN. L. REV. 981 (1996).
351. See RESTATEMENT (THIRD) OF AGENCY § 8.05(2) (2005).
352. Id. § 8.06(1); see also Eugene Volokh, Freedom of Speech and Infor-
mation Privacy: The Troubling Implications of a Right to Stop People from
Speaking About You, 52 STAN. L. REV. 1049, 1058 (2000) (explaining that the
government may impose confidentiality obligations on advisors who do not
“explicitly disclaim any such implicit promise”).
353. Google had $43 billion in advertising revenue for its 2012 fiscal year.
Investor Relations, GOOGLE, http://investor.google.com/financial/tables.html
(last visited Nov. 27, 2013).
2014] SPEECH ENGINES 943
concerns; the fear is that advertisers reject the user’s reality
354
and substitute their own. It is precisely the comprehensive
user profiles that search engines are capable of accumulating in
their ordinary course of operations that makes this outcome so
troubling.
Everything hinges, therefore, on the degree to which users
are aware of the tracking and targeting and are capable of ex-
ercising effective control over them. From a user autonomy per-
spective, the formalistic “consent” of using a website that has a
hyperlink in small type to its privacy terms is a terrible proxy
355
for meaningful choice. A better world would feature what Er-
ic Goldman calls “Coasean filters”: tools that let users and
356
marketers bargain over who receives which messages. But
today’s online world is quite far from that ideal; stronger base-
line protections for user data and an ecology of effective and
usable user-controlled privacy tools will be required to get clos-
er.
C. DEFAMATION
Michael Trkulja would like you to know that he is not a
gangster. He was merely minding his own business having din-
ner in a Melbourne restaurant when a balaclava-clad hit man
357
shot him in the back. The shooting remains unsolved, and a
website named Melbourne Crime posted Trkulja’s picture along
with an article about the case from the Herald Sun Newspa-
358
per. That webpage also had photographs of other notorious
359
criminals and alleged criminals —implying, Trkulja claimed,
360
that he was a member of Melbourne’s criminal underworld.
354. See Chris Jay Hoofnagle et al., Behavioral Advertising: The Offer You
Cannot Refuse, 6 HARV. L. & POL’Y REV. 273, 294 (2012) (“[A]dvertisers do not
see individuals as autonomous beings.”); Tal Zarsky, “Mine Your Own Busi-
ness!”: Making the Case for the Implications of the Data Mining of Personal
Information in the Forum of Public Opinion, 5 YALE J.L. & TECH. 1 (2003).
355. The British videogame chain GameStation takes the cake here: it
added a term to its former website Terms and Conditions that “you agree to
grant Us a non transferable option to claim, for now and for ever more, your
immortal soul.” Marcperton, Read Fine Print or GameStation May Own
Your Soul, CONSUMERIST (Apr. 16, 2010), http://consumerist.com/2010/04/16/
read-fine-print-or-gamestation-may-own-your-soul/ (discussing the immortal
soul term).
356. Goldman, supra note 205, at 1213–18.
357. Trkulja v Google Inc., [2012] VSC 533 ¶ 4 (Austl.).
358. Trkulja v Yahoo! Inc., [2012] VSC 88 ¶¶ 2–4 (Austl.).
359. Id. ¶ 3.
360. Id. ¶ 5.
944 MINNESOTA LAW REVIEW [98:868
Trkulja sued Yahoo! for returning the Melbourne Crime page
361
as a search result for [michael trkulja]. Google had it
even worse: an image search for [michael trkulja] returned
the pictures of actual criminals from the Melbourne Crime
362
page, but captioned with Trkulja’s name.
This time it is the editor theory that argues for liability
and the conduit theory that would exonerate the search engine,
instead of vice versa. The conduit theory treats the search en-
gine as a blameless tool in the service of websites, and therefore
pushes all of the responsibility for content off of the search en-
gine and on to websites. American law, in the form of section
230 of the Communications Decency Act’s immunity for inter-
363
active computer services, adopts the conduit theory. Trkulja’s
only recourse in the United States would be against Melbourne
Crime, not against the search engines that linked to it.
The editor theory, on the other hand, treats the search en-
gine as an active selector and arranger of content. A newspaper
is typically responsible for the material it assembles into each
day’s edition, whether that material came from its own report-
ers, a newswire, advertisers, or another source. So too with a
search engine: it chooses which content to feature and has de-
tailed knowledge about that content. The website will frequent-
ly be unreachable or judgment-proof; the search engine is an
equally culpable but more easily targeted speaker. The Austral-
ian courts followed the editor theory: the Supreme Court of Vic-
364
toria upheld Trkulja’s AU$ 225,000 judgment against Yahoo!
365
and his AU$ 200,000 judgment against Google.
Neither approach is quite correct. We should rather ask
what users want from search in a world where not all infor-
mation is of equal value. A search engine can help by sorting
truth from falsehood—but it can also help simply by helping
361. Id. ¶ 7.
362. Trkulja v Google Inc., [2012] VSC 533 ¶ 2 (Austl).
363. 47 U.S.C. § 230(c)(1) (2006). For cases applying Section 230 to immun-
ize search engines, see Getachew v. Google, Inc., 491 F. App’x 923 (10th Cir.
2012); Neeley v. NameMedia Inc., No. 09-5151, 2011 BL 24617, 97 U.S.P.Q.2d
2029 (W.D. Ark. Jan. 31, 2011); Maughan v. Google Tech., Inc., 143 Cal. App.
4th 1242 (2006); Mmubango v. Google, Inc., No. 12-1300, 2013 WL 664231
(E.D. Pa. Feb. 22, 2013); Parker v. Google, Inc., 422 F. Supp. 2d 492 (E.D. Pa.
2006); Stayart v. Google, Inc., 783 F. Supp. 2d 1055, 1056–57 (E.D. Wis. 2011);
see also Metropolitan Int’l Schools Ltd. v. Designtechnica Corp., [2011] 1 WLR
1743, [2009] EWHC 1765 (Q.B.) (finding Google not liable for allegedly defam-
atory statements in excerpt from website).
364. Trkulja v Yahoo! Inc., [2012] VSC 88 ¶ 60 (Austl).
365. Trkulja v Google Inc., [2012] VSC 533 ¶ 55 (Austl).
2014] SPEECH ENGINES 945
users find relevant information on a topic. This latter function
is more basic: one cannot reliably draw accurate conclusions
without access to the full range of data on a topic. When a
search engine performs the latter role for its users—telling
them what others have said—it does so without endorsing the
truth of the content it excerpts or links to. A ranking is a guess
that the user will find the content relevant, nothing more. Like
a newspaper reporting on the controversy over public officials’
366
defamatory statements, a search engine performs a valuable
service by telling its users about the existence of a debate in the
367
first place.
That is, the advisor theory reminds us that what is truly at
stake is users’ access to information. A decision that certain
content ought not to be indexed—because it is defamatory, be-
cause it is harassing, because it incites racial hatred, or be-
cause it will inevitably cause moral rot, tooth decay, and alien
invasion—should be recognized for what it is: a decision by
government to censor the information available to search users.
Of course, there is a good reason for this censorship: defama-
tion law reflects a collective judgment that harmful lies about
people ought not be repeated. But the collateral consequences
of a duty on search engines to avoid defamatory results are
368
likely to be especially severe. The crucial facts—whether the
complained-of statements are true or false—are not typically
likely to be in the possession of the search engine. And the sub-
tle shades of meaning involved in parsing allegedly defamatory
statements make even the notoriously difficult task of as-
sessing fair use seem simple by comparison. Search engines
need clear and well-sheltered safe harbors from defamation lia-
bility.
But section 230 goes too far by providing search engines an
absolute immunity for content supplied by websites, regardless
369
of knowledge or intent. If Trkulja has sued Melbourne Crime
366. See Ashley Messenger, The Problem with New York Times, Co. v. Sul-
livan: An Argument for Moving from a “Falsity Model” of Libel Law to a
“Speech Act Model,” 11 FIRST AMEND. L. REV. 172, 177 (2012).
367. See Nieman v. Versuslaw, Inc., No. 3:12-cv-3104, 2012 BL 196699, 40
Med. L. Rptr. 2191 (C.D. Ill. Aug. 03, 2012) (holding that the First Amend-
ment bars claims against search engines for linking to court documents).
368. See Felix Wu, Collateral Censorship and the Limits of Intermediary
Immunity, 87 NOTRE DAME L. REV. 293 (2011) (discussing § 230 and the collat-
eral censorship problem); Grimmelmann, supra note 63 (arguing that the
problem will be especially severe for search engines).
369. For an overview of proposals to modify section 230, see JOEL
946 MINNESOTA LAW REVIEW [98:868
and won, the argument for leaving Yahoo! and Google entirely
alone is much weaker. The same is true if Melbourne Crime is
unreachable because it is overseas or anonymous, or if the
same false claims are repeated on so many websites that suing
them individually is obviously infeasible. Search engines’ im-
munity should be limited when victims supply sufficiently sub-
370
stantiated proof that the linked-to material is defamatory.
And some search engines don’t even deserve this much protec-
tion. Imagine a search engine called the Scandal Rag that re-
sponds to every search query by linking to third-party page ac-
cusing Trkulja of murder and arson without a shred of proof.
The Scandal Rag substitutes its own agenda for users’ goals; it
has stepped out of the kind of role for which immunity makes
sense. It is acting like a publisher, rather than an advisor, and
371
the law should treat it as one.
372
Censorship is one thing; secret censorship quite another.
When the law requires search engines not to link to certain
content, the very least it owes to users is an explanation.
Google provides ready examples of what to do. When the Chi-
nese government required it not to return certain search re-
sults relating to the Tiananmen Square crackdown or to Falun
Gong, Google decorated the Google.cn results pages that would
have contained those links with a disclaimer warning that
373
some results had been removed to comply with local laws.
REIDENBERG ET AL., FORDHAM CTR. ON LAW & INFO. POLICY, SECTION 230 OF
THE COMMUNICATIONS DECENCY ACT: A SURVEY OF THE LEGAL LITERATURE
AND REFORM PROPOSALS (2012).
370. In copyright, section 512 of the Digital Milennium Copyright Act lim-
its search engines’ immunity in just this way. See 17 U.S.C. § 512(d)(3) (2012).
But the threshold there is too low, because courts have interpreted the provi-
sion prohibiting bad-faith notices so narrowly that copyright owners can file
takedown notices, even in cases of obvious fair use, with impunity. See Rossi v.
Motion Picture Ass’n of Am., 391 F.3d 1000, 1004–05 (9th Cir. 2004) (inter-
preting Section 512 as setting a “subjective” standard, so that even unreason-
able beliefs about infringement can be the basis for takedown notices). A bet-
ter provision, in copyright and in defamation, would require that the
complainant provide more evidence in the notice, and impose greater penalties
for bad-faith or unreasonable notices.
371. Section 512(d) is a better model here. See 17 U.S.C. § 512(d)(1)(A)
(2012) (denying safe harbor when information location tool has “actual
knowledge that the material or activity is infringing”).
372. Cf. Derek E. Bambauer, Cybersieves, 59 DUKE L.J. 377, 393 (2009) (ar-
guing that Internet filtering is more legitimate when countries are transpar-
ent about what material is blocked and why).
373. See Grimmelmann, supra note 97, at 947–50; see also Alan Eustace,
Better Search in Mainland China, GOOGLE INSIDE SEARCH (May 31, 2012,
9:00 AM), http://insidesearch.blogspot.com/2012/05/better-search-in-mainland
2014] SPEECH ENGINES 947
And when Google receives copyright takedown notices under
section 512(d) of the Digital Millennium Copyright Act
(DMCA), it forwards them to the Chilling Effects clearinghouse
374
to document the resulting removals. But Google also provides
ready examples of what not to do. Starting in August 2012, it
added a new signal, downgrading sites that received high num-
bers of DMCA takedown requests—even for content that had
375
not been the subject of a DMCA notice. This move is neither
376
required by copyright law nor relevance-enhancing from us-
377
ers’ perspective. If Google hides webpages it thinks users are
looking for, it should be honest with them and say on the
search results page that it has done so.
D. TRADEMARK
According to Google, [rosetta stone] has many mean-
ings. It refers to the Rosetta Stone, the Egyptian stele that
made it possible to decipher hieroglyphs. It refers to the well-
known line of language-learning software identified by the
ROSETTA STONE trademark. And it refers to a wide range of
online sites where one can buy language-learning software—
some of it authorized ROSETTA STONE software, some of it
378
not. This last category is responsible for all of the trouble.
There is a long-running battle between trademark owners
and search engines over keyword advertising—which supplies
the money that keeps Google and its competitors in the search
379
business at all. The trademark owners hate it when competi-
-china.html (describing new notices to warn Chinese users when a search term
they are typing might cause their Internet connection to be interrupted).
374. See FAQ, GOOGLE TRANSPARENCY REPORT, http://www.google.com/
transparencyreport/removals/copyright/faq/ (last visited Nov. 27, 2013) (“We
link in our search results to the requests published by Chilling Effects in place
of removed content when we are able to do so legally.”).
375. Singhal, supra note 198.
376. See generally UMG Recordings, Inc. v. Shelter Capital Partners, 718
F.3d 1006 (9th Cir. 2013).
377. Eric Goldman, Why Did Google Flip-Flop on Cracking Down on
“Rogue” Websites? Some Troubling Possibilities, TECH. & MKTG. LAW BLOG
(Aug. 22, 2012, 9:46 AM), http://blog.ericgoldman.org/archives/2012/08/
why_did_google.htm.
378. See Rosetta Stone Ltd. v. Google, Inc., 676 F.3d 144, 172–73 (2012)
(allowing Rosetta Stone’s trademark claims against Google to proceed). The
case subsequently settled on undisclosed terms. See Stipulation of Voluntary
Dismissal with Prejudice, Rosetta Stone Ltd. v. Google, Inc., 676 F.3d 144 (4th
Cir. 2012) (No. 1:09-cv-00736).
379. The academic literature on keyword advertising and trademarks
online is immense. Some helpful sources include Margreth Barrett, Internet
948 MINNESOTA LAW REVIEW [98:868
tors use their trademarks as keywords to trigger advertise-
ments; they have regularly sued both the competitors and
380
search engines, the former with somewhat more success. As
against search engines, the consensus seems to be that yes,
such uses are potentially infringing, but no court has entered a
judgment that a search engine was actually infringing because
381
of its keyword advertising.
Both the conduit and editor theories are ambiguous here.
On the conduit theory, perhaps the trademark confers an ex-
clusive right in the trademark owner to use the mark to attract
customers, so any diversion of customers looking for the mark
owner is a misdirection of traffic. A search for [coke] should
382
lead to the real thing, not an ad for Mocha-Cola. Or perhaps
the search engine is merely a conduit for advertisers’ messages,
and does not take responsibility for them. Coke should have it
out with the makers of Mocha-Cola, not with Google.
The editor theory is no better. One could argue that the
search engine is not a merchant supplying goods and services
to compete with trademark owners; it is merely arranging in-
formation about websites in a convenient form, like a drugstore
383
placing every brand of cola in the same section. Or one could
argue that the search engine is crafting a deceptive message for
users: it was asked for [coke] but it served up Mocha-Cola ads
instead.
The advisor theory returns our attention to users. They are
the ones who create the many meanings of [rosetta stone];
the search query is always an approximation of their actual in-
tentions. Some want to buy Rosetta Stone software and want
the official site, or a retailer settling it, or a comparison of pric-
Trademark Suits and the Demise of “Trademark Use”, 39 U.C. DAVIS L. REV.
371 (2006); Graeme B. Dinwoodie & Mark D. Janis, Confusion over Use:
Contextualism in Trademark Law, 92 IOWA L. REV. 1597 (2007); Stacey L.
Dogan & Mark A. Lemley, Trademarks and Consumer Search Costs on the In-
ternet, 41 HOUS. L. REV. 777 (2004); Eric Goldman, Deregulating Relevancy in
Internet Trademark Law, 54 EMORY L.J. 507 (2005); Greg Lastowka, Google’s
Law, 73 BROOK. L. REV. 1327 (2008).
380. See, e.g., 1-800 Contacts, Inc. v. Lens.Com, Inc., 722 F.3d 1229, 1237
(10th Cir. 2013).
381. For an example of an especially definitive win for Google, see Google
Inc. v Austl. Competition & Consumer Comm’n, [2013] HCA 1 (Austl.).
382. See, e.g., Coca-Cola Co. v. Overland, Inc., 692 F.2d 1250, 1252 (9th Cir.
1982) (holding that it was trademark infringement to serve Pepsi to customer
who ordered “Coke” without disclosing substitution to customer).
383. See Eric Goldman, Brand Spillovers, 22 HARV. J.L. & TECH. 381, 382–
83 (2009).
2014] SPEECH ENGINES 949
es across multiple retailers. Some are engaged in product re-
search: they want user and expert reviews; others are looking
to learn more about Rosetta Stone’s competition, using the
name of the category’s best-known brand as a rough synonym
for [language learning software]. And some really are
just looking for the trilingual stele.
This ambiguity means that there is substantial danger in
giving any one website exclusive rights to control a search que-
ry: it allows the website to divert a wide range of users with di-
verse interests. At the same time, a disloyal search engine can
steer users wrong by taking money to show them ads intended
to divert them from the websites they’re actually looking for.
Striking the right balance is a subtle affair. It is easy for a
smartphone shopper to glance at and reject one ad, or to hit the
back button when she realizes these aren’t the [droids] she’s
looking for. But the combined effect of a dozen such ads, or a
hundred, can be significant. At some point, the sheer clutter
makes it impossible for the user to find the smartphones she
seeks. And, of course, an openly deceptive ad coupled with a de-
ceptive website can indeed trick the purchaser into buying the
wrong thing.
Thus, the advisor theory leads us naturally back to the
question trademark law is also supposed to ask: are consumers
likely to be confused? This is a fact-sensitive inquiry; it depends
on users’ reasons for using a particular query, on how results
are presented, and on how clearly the paid nature of keyword
384
ads is disclosed. One recent case explained that even when
consumers searched for the plaintiff’s trademark on Amazon
and the search results did not contain any of plaintiff’s goods,
confusion was unlikely:
Additionally, the instant situation does not appear to be a case of
palming off in the traditional sense. It is akin to the consumer asking
for a Coca-Cola and receiving a tray with unopened, labeled, authen-
tic cans of Pepsi-Cola, RC Cola, Blue Sky Cola, Dr. Pepper, and
Sprecher Root Beer, and a copy of Coca Kola: The Baddest Chick, by
Nisa Santiago. This is a substitution, but given the context it is not
385
infringing because it is not likely to confuse.
The most recent and careful empirical study in the area
384. Compare Network Automation, Inc. v. Advanced Sys. Concepts, Inc.,
638 F.3d 1137, 1154 (9th Cir. 2011) (paid placement labeled as such), with
Playboy Enters., Inc. v. Netscape Commc’ns, 354 F.3d 1020, 1030 (9th Cir.
2004) (paid placement not labeled as such).
385. Multi Time Machine, Inc. v. Amazon.com, 926 F. Supp. 2d 1130, 1137
n.4 (C.D. Cal. 2013).
950 MINNESOTA LAW REVIEW [98:868
found “little evidence of . . . consumer confusion regarding the
source of goods, but only a small minority of consumers correct-
ly and consistently distinguished paid ads from unpaid search
386
results.” These results suggest that Google ought to prevail in
other trademark keyword cases—but also that regulators
should require clearer differentiation between unpaid organic
387
search results and paid search advertisements.
CONCLUSION
A good search engine advises its users, helping them to be-
come active listeners, and enabling them to act autonomously.
Each of these points opens up promising avenues for further
inquiry.
First, there is the application of the advisor theory to other
problems in search-engine law. This Article has dealt primarily
with search bias, arguing that a search result is problematical-
ly “biased” when it constitutes bad-faith advice to the user. The
Article has also given brief attention to problems of copyright,
privacy, defamation, and trademark. But the advisor theory
can also provide insights into the antitrust cases against
Google, into search engines’ obligation to filter copyright-
infringing results, into search engines’ obligations in dealing
with repressive authoritarian governments, into the role of
search in open access to government information, and into the
problem of web spam targeting search engines, among other is-
sues.
Second, there is active listening. Search engines are an ob-
vious case of active listening—but far from the only one. De-
386. David J. Franklyn & David A. Hyman, Trademarks as Search Engine
Keywords: Much Ado About Something?, 26 HARV. J.L. & TECH. 418, 484
(2013).
387. The FTC currently advises search engines to engage in “clear and con-
spicuous disclosures” of the sponsorship of search ads. Letter from Heather
Hippsley, Acting Assoc. Dir., FTC Div. of Adver. Practices, to Gary Ruskin,
Exec. Dir. of Commercial Alert (June 27, 2002), available at http://www.ftc
.gov/os/closings/staff/commercialalertletter.shtm; Sample Search Engine Gen-
eral Letter from Mary Engle, Assoc. Dir. for Adver. Practices, Fed. Trade
Comm’n (June 24, 2013), http://www.ftc.gov/os/2013/06/130625searchengine
generalletter.pdf. It is unclear whether simple labels on search results do an
adequate job of informing users. See David J. Franklyn & David A. Hyman,
Review of the Likely Effects of Google’s Proposed Commitments Dated April 25,
2013, FAIR SEARCH EUROPE (July 1, 2013), http://www.fairsearcheurope.eu/wp
-content/uploads/2013/07/FS-Survey-Results-on-Effects-of-Googles-Proposed-
Commitments.pdf (finding little effect on consumer behavior from, and sub-
stantial consumer misunderstand of, proposed labels of Google-affiliated con-
tent).
2014] SPEECH ENGINES 951
scriptively, the fact that listeners can and do make choices
about which speech to receive helps explain numerous First
388
Amendment doctrines. And normatively, empowering listen-
ers to make effective choices among speakers is a worthy
389
goal. A well-developed theory of active listening has the po-
tential to enrich First Amendment theory and doctrine. It could
yield insights into the captive audience doctrine, the status of
commercial speech, Internet filtering, targeted advertising, tel-
ecommunications regulation, and anonymous speech, among
other topics.
And third, user autonomy is an important principle in
computer and Internet law, one with implications well beyond
search engines. Consider, for example, the problem of malware.
Modern operating systems make it difficult or impossible to in-
stall unknown and untrusted software. These rules restrict us-
ers’ choices about which software to run—but might they also
enhance users’ effective autonomy by protecting them from
malware that disables their computers and spies on their
online activity? Other issues that could benefit from a more
systematic focus on user autonomy include digital rights man-
agement, unauthorized access to computer systems, online con-
tracting, ad-blocking software, Do Not Track, and cell-phone
unlocking.
As for the advisor theory itself, this way of thinking about
search may seem cynical about the motivations of websites and
search engines. Websites are clamoring to be found; they will
attempt to trick search engines into ranking them highly; fail-
ing that, they will turn to the government and demand the
same. Search engines, for their part, have the means to mislead
users. Where their commercial interests are at stake, they can
be expected to put those interests first if they expect to be able
to get away with it. In any case pitting a website against a
search engine, it is best to read the briefs with a grain of salt in
each hand.
388. See, e.g., Sable Commc’ns of Cal., Inc., v. FCC, 492 U.S. 115, 128
(1989) (“[T]he message received by one who places a call to a dial-a-porn ser-
vice is not so invasive or surprising that it prevents an unwilling listener from
avoiding exposure to it.”).
389. See, e.g., Rowan v. U.S. Post Office Dep’t, 397 U.S. 728, 736 (1970)
(“[A] sufficient measure of individual autonomy must survive to permit every
householder to exercise control over unwanted mail.”); id. at 738 (“If this pro-
hibition operates to impede the flow of even valid ideas, the answer is that no
one has a right to press even ‘good’ ideas on an unwilling recipient.”).
952 MINNESOTA LAW REVIEW [98:868
But cynicism should not be mistaken for pessimism. The
story that the advisor theory tells is profoundly hopeful. It is
hopeful about users’ capacity for self-fulfillment, and it is hope-
ful about what better search will do for us all. Search is worth
getting right because it matters, and will continue to matter as
long as humans are still asking questions of the world and of
390
each other.
390. Cf. VAIDHYANATHAN, supra note 176, at 209–10 (“If we want to create
a vital global public sphere for the digital era by offering the best and the most
information to the largest number of people around the world . . . [w]e can’t
just hope that some big, rich company will do it for us. That’s simply irrespon-
sible. . . . The future of knowledge—and thus the future of the species—
depends on getting this right.”).