Gonzalez v. Google: The Case for Protecting "Targeted Recommendations"

Duke Law Journal Online

VOLUME 72 APRIL 2023

GONZALEZ V. GOOGLE: THE CASE FOR

PROTECTING “TARGETED

RECOMMENDATIONS”

TOMER KENNETH† & IRA RUBINSTEIN††

BSTRACT

Does Section 230 of the Communications Decency Act protect

online platforms (e.g., Facebook, YouTube, and Twitter) when they

use recommendation algorithms? Lower courts upheld platforms’

immunity, notwithstanding notable dissenting opinions. The Supreme

Court considers this question in Gonzalez v Google, LLC. Plaintiffs

invite the Court to analyze “targeted recommendations” generically

and to revoke Section 230 immunity for all recommended content. We

think this would be a mistake.

This Article contributes to existing scholarship about Section 230

and online speech governance by adding much needed clarity to the

desirable—and undesirable—regulation of recommendation

algorithms. Specifically, this Article explains the technology behind

algorithmic recommendations, the questions it raises for Section 230

immunity, and the stakes in Gonzalez. It opposes generically revoking

Section 230 immunity for all uses of recommendation algorithms.

Instead, it illustrates and defends a nuanced approach for the desired

outcome of Gonzalez and for future possible regulation of

recommendation algorithms.

† JSD Candidate; Fellow, Information Law Institute, New York University School of Law.

†† Senior Fellow, Information Law Institute, New York University School of Law. We

thank Neli Frost, Eugene Volokh, and the fellows at NYU’s Information Law Institute for helpful

discussions and comments.

2023] GONZALEZ V. GOOGLE 177

NTRODUCTION

Are online platforms (e.g., Facebook, YouTube, and Twitter)

legally immune when they use recommendation algorithms to match

specific content to specific users? This is the question now before the

Supreme Court in Gonzalez v. Google, LLC.

Although lower courts

have upheld immunity,

a few notable dissenting opinions have

rejected this conclusion, arguing for limited application of Section 230

in those contexts.

Granting certiorari in Gonzalez and framing the

question at issue in very broad terms, the Supreme Court seems poised

to reach sweeping conclusions about the application of Section 230

immunities to recommendation algorithms.

This Article takes issue with the Supreme Court’s framing and

argues that a generic application of Section 230 to recommendation

algorithms is a mistake. The Article defends the majority opinion in

the lower courts and argues that there are better ways to address online

speech and regulate online platforms. The Article complements

existing scholarship by discussing the often-ignored relationship

between Section 230 and recommendation algorithms. This

contribution matters for Gonzalez but also for legislative efforts to

amend Section 230

and for other attempts to regulate

recommendation algorithms.

In Part I, we briefly present the relevant background: the

technology behind algorithmic recommendation, the context of

Section 230, and the question in Gonzalez. Then, in Part II, we criticize

the Supreme Court’s framing of the relationship between Section 230

and recommendation algorithms as too generic. This broad framing

invites an overinclusive analysis, misunderstands the relevant

technology, and forces a false dilemma: either undermining the

1. Gonzalez v. Google LLC, 2 F.4th 871 (9th Cir. 2021), cert. granted, No. 21-1333, 2022

WL 4651229 (U.S. Oct. 3, 2022), cert. granted sub nom. Twitter, Inc. v. Taamneh, No. 21-1496,

2022 WL 4651263 (U.S. Oct. 3, 2022).

2. Force v. Facebook, Inc., 934 F.3d 53, 71 (2d Cir. 2019); Gonzalez, 2 F.4th at 913.

3. See, e.g., Force, 934 F.3d at 76 (Katzmann, C.J., concurring); Gonzalez, 2 F.4th at 918

(Gould, J., concurring).

4. See infra Part II.

5. See, e.g., J

ASON A. GALLO & CLARE Y. CHO, CONG. RSCH. SERV., R46662, SOCIAL

MEDIA: MISINFORMATION AND CONTENT MODERATION ISSUES FOR CONGRESS 6–8 (2021).

6. Compare NetChoice, LLC v. Paxton, 573 F.Supp 3d 1092, 1099 (W.D. Tex. 2022), rev’d,

49 F.4th 439 (5th Cir. 2022) (granting a preliminary injunction that barred enforcement of a Texas

social media law restricting content moderation), with NetChoice, LLC v. Att’y Gen., 34 F.4th

1196, 1232 (11th Cir. 2022) (upholding several provisions in a Florida social media law regulating

content moderation).

178 DUKE LAW JOURNAL ONLINE [Vol. 72:176

protections that helped make online platforms so desirable for users or

not regulating online platforms at all.

In Part III, we argue that the Supreme Court should uphold the

Ninth Circuit majority’s view in Gonzalez. According to this view,

platforms forfeit Section 230 protections only if they make material

contributions to the content that users upload.

We explain the

advantages of this application-specific approach for Section 230

generally and for recommendation algorithms more specifically. We

also consider the shortcomings of the dissenting opinions, which would

exclude algorithmic recommendations from Section 230 immunity.

One such view sees all uses of recommendation algorithms as

conveying a message; another excludes only recommending

connections to other users, groups, or pages.

Both should be avoided.

Admittedly, our position offers little recourse for many of the

perverse outcomes of the prior interpretations of Section 230.

share many of these concerns and believe that governments can and

should do more to rein in online platforms and to cultivate a better

online speech environment. However, we think that excluding

platforms’ use of recommendation algorithms from Section 230

immunities is the wrong approach. In the concluding section, we point

to more desirable solutions, such as carving out narrow exceptions to

Section 230 or amending the statute to ensure that firms engage in

Good Samaritan screening as a condition of immunity. We also briefly

consider requiring the use of technological friction to mitigate

algorithmic amplification or using soft regulation that provides

guidance to online platforms.

ON SECTION 230 AND RECOMMENDATION ALGORITHMS

There is an abundance of scholarly writing on the historical

background and genealogy of Section 230 of the Communications

Decency Act of 1996.

For our purposes, a brief introduction suffices.

Section 230 provides “interactive computer services” immunity from

7. See infra note 14; Part III.A.

8. See infra Part III.B–C.

9. See Danielle Keats Citron & Benjamin Wittes, The Internet Will Not Break: Denying

Bad Samaritans § 230 Immunity, 86 F

ORDHAM L. REV. 401, 401–03 (2017) (summarizing potential

harms to minors and young adults under broad Section 230 immunity).

10. For some examples, see generally Adam Candeub, Reading Section 230 As Written, 1 J.

FREE SPEECH L. 139 (2021); Eric Goldman, An Overview of the United States’ Section 230 Internet

Immunity, in O

XFORD HANDBOOK OF ONLINE INTERMEDIARY LIABILITY 154 (Giancarlo Frosio

ed., 2020); Danielle Keats Citron & Benjamin Wittes, The Problem Isn’t Just Backpage: Revising

Section 230 Immunity, 2 G

EO. L. TECH. REV. 453 (2018).

2023] GONZALEZ V. GOOGLE 179

(1) liability as “publisher or speaker of any information” that a third-

party uploads and from (2) civil liability for the removal of content

under certain circumstances.

Congress hoped that Section 230 would

promote the continued development of the Internet and online services

and preserve its vibrancy as an educational and informational resource

for all citizens while also encouraging the removal of offensive content

without exposing these services to publisher’s liability.

Courts

adopted a broad view of Section 230. This included, first, interpreting

“interactive computer services” as covering new social media platforms

like Facebook and Twitter,

emphasizing that statutory immunity

protected these services against liability for “any information” that

third parties published,

and, second, imposing liability only when

platforms make a “material contribution” to the content uploaded by

users.

Section 230 has succeeded in its main goal: facilitating the creation

of a vibrant social networking environment online, led and governed

by private companies.

However, in recent years, scholars,

11. 47 U.S.C. § 230(c); see also VALERIE C. BRANNON & ERIC N. HOLMES, CONG. RSCH.

SERV., R46751, SECTION 230: AN OVERVIEW 1 (2021) (“[Section 230] sought to allow users and

providers of ‘interactive computer services’ to make their own content moderation decisions,

while still permitting liability in certain limited contexts.”).

12. 47 U.S.C. § 230(a)–(b).

13. See, e.g., Klayman v. Zuckerberg, 753 F.3d 1354, 1355 (D.C. Cir. 2014) (classifying

Facebook as an interactive computer service); Fields v. Twitter, Inc., 217 F. Supp. 3d 1116, 1118

(N.D. Cal. 2016) (classifying Twitter as an interactive computer service).

14. See, e.g., Gonzalez v. Google LLC, 2 F.4th 871, 886–87, 896 (9th Cir. 2021), cert. granted,

No. 21-1333, 2022 WL 4651229 (U.S. Oct. 3, 2022), cert. granted sub nom. Twitter, Inc. v.

Taamneh, No. 21-1496, 2022 WL 4651263 (U.S. Oct. 3, 2022) (emphasizing that Congress made a

policy decision to provide broad protection under Section 230, protecting any information);

Carafano v. Metrosplash.com., Inc., 339 F.3d 1119, 1122–25 (9th Cir. 2003) (same); Doe v. Internet

Brands, Inc., 824 F.3d 846, 851–54 (9th Cir. 2016) (same).

15. See Fair Hous. Council of San Fernando Valley v. Roommates.com, LLC, 521 F.3d 1157,

1169–71 (9th Cir. 2008) (distinguishing “providing neutral tools” from “materially contributing”

to the alleged unlawfulness); Marshall’s Locksmith Serv. v. Google, LLC, 925 F.3d 1263, 1271

(D.C. Cir. 2019) (holding that algorithms that convert third-party input of location into picture

form use neutral means and therefore enjoy Section 230 immunity).

16. 47 U.S.C. § 230(b)–(c); Danielle Keats Citron, How To Fix Section 230, B.U.

L. REV.

(forthcoming) (manuscript at 3) (on file with authors) (“The absence of liability meant that search

engines could link to sites, blogs, and other online activity without fear that they would be liable

for defamatory comments.”).

17. See generally, e.g., Gautam Hans, Revisiting Roommates.com, 36 B

ERKELEY TECH. L.J.

1228 (2022) (evaluating potential reforms of Section 230 that would further the goals of civil rights

protections); Olivier Sylvain, Platform Realism, Informational Inequality, and Section 230 Reform,

131 Y

ALE L.J.F. 475 (2021) (arguing in favor of Section 230 reform that would result in outcomes

more consistent with settled consumer-protection and civil-rights laws); Danielle Keats Citron &

Mary Anne Franks, The Internet as a Speech Machine and Other Myths Confounding Section 230

180 DUKE LAW JOURNAL ONLINE [Vol. 72:176

legislators,

and courts

have questioned the breadth of Section 230

immunity. The major concern is that Section 230 grants platforms too

much discretion and power to govern the ever-growing aspects of

online life.

In Gonzalez v. Google, LLC, petitioners seek to limit the scope of

Section 230 immunity.

The petitioners, plaintiffs in a Ninth Circuit

case, are relatives of victims of terrorist attacks for which ISIS claimed

responsibility.

They sought to establish the platforms’ liability under

the Anti-Terrorism Act (ATA) for content that circulated on those

platforms.

Recognizing that a broad interpretation of Section 230

immunity protects platforms from liability for terrorist content

uploaded by third-party users,

plaintiffs opted to distinguish their

claims by arguing that platforms’ use of recommendation algorithms is

not protected under Section 230.

Allegedly, platforms

“‘recommended ISIS videos to users’ and enabled users to ‘locate other

videos and accounts related to ISIS,’ thereby assisting ISIS in spreading

its message.”

Conversely, the platforms argued (among other things)

that Section 230 protects their use of algorithms to recommend specific

content to specific users.

Reform, 2020 U. CHI. LEGAL F. 45 (recommending changes to Section 230 that would condition

immunity on reasonable moderation practices).

18. G

ALLO & CHO, supra note 5, at tbl.B-1 (listing over two dozen Section 230 reform

proposals introduced in the 116th Congress).

19. See, e.g., Biden v. Knight First Amend. Inst. at Columbia Univ., 593 U.S. 1220, 1221

(2021) (Thomas, J., concurring) (criticizing the immense power that private platforms have over

online speech and the need to regulate them); Gonzalez, 2 F.4th at 912–13, 923 (“Whether social

media companies should continue to enjoy immunity for the third-party content they publish, and

whether their use of algorithms ought to be regulated, are pressing questions that Congress should

address.”).

20. Gonzalez, 2 F.4th at 886.

21. Id. at 880–85.

22. Id. at 880. The ATA allows U.S. nationals to recover damages for injuries suffered “by

reason of an act of international terrorism,” 18 U.S.C. § 2333(a), and extends liability to “any

person who aids and abets, by knowingly providing substantial assistance” to a person who

commits an act of international terrorism, 18 U.S.C. § 2333(d).

23. Force v. Facebook, Inc., 934 F.3d 53, 65–66 (2d Cir. 2019).

24. Gonzalez, 2 F.4th at 881, 894–95.

25. Id. at 881. Plaintiffs in the other two cases decided in Gonzalez—Taamneh and

Clayborn—make roughly similar claims, alleging that YouTube, Facebook and Twitter failed to

do enough to stop ISIS from using their platforms to promote its messages and recruit terrorists.

Id. at 883–84.

26. Id. at 882, 894–95.

2023] GONZALEZ V. GOOGLE 181

To assess these claims, we need to better understand the

technology of recommendation algorithms.

Social media platforms

process copious amounts of user-generated content. Given the scale

and variation of content involved, platforms rely on algorithmic

automation to manage content with the goals of making the platforms

interesting and maximizing user engagement.

The two main ways to

algorithmically manage content are content moderation and

algorithmic recommendation.

Content moderation means, roughly, (1) fitting content into

predefined categories based on published “community guidelines” and

(2) issuing warnings about, demoting, or removing content that violates

these guidelines.

For instance, a platform like Twitter relies on

content moderation algorithms to identify uploaded content as

“COVID-19 misinformation” and enforce (or cease to enforce) its

policy of removing “demonstrably false or potentially misleading

content that has the highest risk of causing harm.”

Conversely, algorithmic recommendation optimizes the use of

(permissible) content on the platform.

Platforms use

recommendation algorithms to rank content algorithmically and, based

on these rankings, to promote specific content to particular users or

distribute certain content more broadly. While content moderation

asks how best to classify content, recommendation algorithms ask how

best to use this content in order to maximize desired outcomes

(typically, user engagement). Social media firms collect and analyze

hundreds (or even thousands) of data points and feed this data to

recommendation algorithms designed to predict what specific content

will keep specific users most engaged. For example, Facebook’s

newsfeed algorithm relies on predictive models that learn what drives

users to interact with a piece of content “based on who posted it, what[]

27. See generally Ira S. Rubinstein & Tomer Kenneth, Taming Online Public Health

Misinformation, 60 H

ARV. J. ON LEGIS. (forthcoming 2023), https://ssrn.com/abstract_id=4192903

[https://perma.cc/F97T-VTM3] (discussing the technological background of algorithmic

recommendation and content moderation).

28. See generally, e.g., T

ARLETON GILLESPIE, CUSTODIANS OF THE INTERNET:

PLATFORMS, CONTENT MODERATION, AND THE HIDDEN DECISIONS THAT SHAPE SOCIAL

MEDIA (2018) (viewing content moderation as a fundamental aspect of social media platforms

and suggesting that algorithmic choice of content is what draws users in and keeps them on a given

platform).

29. Rubinstein & Kenneth, supra note 27, at 52–56.

30. Natasha Lomas, Twitter Says It’s No Longer Enforcing COVID-19 Misleading

Information Policy, T

ECHCRUNCH (Nov. 29, 2022), https://techcrunch.com/2022/11/29/twitter-co

vid-29-misleading-info-policy-change [https://perma.cc/ZAA3-4WCE].

31. Rubinstein & Kenneth, supra note 27, at 52–56.

182 DUKE LAW JOURNAL ONLINE [Vol. 72:176

it[‘s] about, whether it contains an image, or a video, what’s in the

video, how recent it is, how many of our friends liked or shared it and

so on.”

YouTube and other platforms follow a similar approach.

In sum, there are important technological differences between

content moderation and algorithmic recommendation. These

differences have legal implications. In previous writing, we have argued

that content moderation regulations are content-based and hence

subject to strict scrutiny under the First Amendment, while

recommendation algorithms are content-neutral and hence should

receive intermediate scrutiny.

In a similar vein, the plaintiffs in

Gonzalez seek to distinguish the treatment of recommendation

algorithms from content moderation for the purposes of Section

230(c)(2) protections. Should they prevail? Do the technological

distinctions between content moderation and algorithmic

recommendation warrant an exclusion of Section 230 protections for a

platform’s uses of recommendation algorithms? Are online platforms

(e.g., Facebook, YouTube, and Twitter) legally immune when they

recommend specific content to specific users?

Most courts have held that Section 230 protects platforms in using

recommendation algorithms.

We agree with this conclusion and find

32. See, e.g., SINAN ARAL, THE HYPE MACHINE 84 (2020); see also TANIA BUCHER,

IF . . . THEN: ALGORITHMIC POWER AND POLITICS 78 (2018) (identifying similar factors

Facebook considers in determining the “relevancy score” of posts in a use’s newsfeed);

Akos

Lada, Meihong Wang & Tak Yan, How Machine Learning Powers Facebook’s News Feed

Ranking Algorithm, E

NGINEERING AT META (Jan. 26, 2021), https://engineering.fb.com/2021/01/

26/ml-applications/news-feed-ranking [https://perma.cc/VGL3-B6EV] (describing the technical

aspects of Facebook’s ranking algorithm).

33. See, e.g., M

OZILLA FOUNDATION, YOUTUBE REGRETS 13–14 (2019), https://assets.mof

oprod.net/network/documents/Mozilla_YouTube_Regrets_Report.pdf [https://perma.cc/L4HR-

UQBW].

34. See Rubinstein & Kenneth, supra note 27, at 56. For an opposing view, see generally

Daphne Keller, Amplification and Its Discontents: Why Regulating the Reach of Online Content

Is Hard, 1 J.

FREE SPEECH L. 227 (2021). Recent court rulings diverge over how to analyze content

moderation for First Amendment purposes. Compare NetChoice, LLC v. Paxton, 573 F. Supp. 3d

1092, 1099 (W.D. Tex. 2022), rev’d, 49 F.4th 439 (5th Cir. 2022) (granting a preliminary injunction

that barred enforcement of a Texas social media law restricting content moderation), with

NetChoice, LLC v. Att’y Gen., 34 F.4th 1196, 1232 (11th Cir. 2022) (upholding several provisions

in a Florida social media law regulating content moderation).

35. See Gonzalez v. Google LLC, 2 F.4th 871, 894 (9th Cir. 2021), cert. granted, No. 21-1333,

2022 WL 4651229 (U.S. Oct. 3, 2022), cert. granted sub nom. Twitter, Inc. v. Taamneh, No. 21-

1496, 2022 WL 4651263 (U.S. Oct. 3, 2022) (“Though we accept . . . that Google’s algorithms

recommend ISIS content to users, the algorithms do not treat ISIS-created content differently

than any other third-party created content, and thus are entitled to § 230 immunity.”); Force v.

Facebook, Inc., 934 F.3d 53, 70 (2d Cir. 2019) (“The algorithms take the information provided by

Facebook users and ‘match’ it to other users . . . based on objective factors applicable to any

content . . . . [This use of recommendation algorithms] is not enough to hold Facebook

2023] GONZALEZ V. GOOGLE 183

the dissenting opinions’ reasoning unpersuasive.

But before turning

to this discussion, the next Part takes up more pressing matters—the

Supreme Court’s pending decision about recommendation algorithms.

II.

THE CHALLENGE OF GONZALEZ V. GOOGLE, LLC

The Supreme Court is set to address the application of Section 230

to uses of recommendation algorithms in Gonzalez v. Google, LLC.

This section raises questions about the Court’s decision to take up the

case and argues that the Court’s apparent rationale for considering it

is ill-advised. Granting certiorari in Gonzalez and commenting in other

cases, the Supreme Court has shown an inclination to exclude

“targeted recommendations” from Section 230 protection altogether.

We take issue with this view. We think it is a mistake to analyze the use

of recommendation algorithms in such broad strokes. And, we think

that, in most cases, the use of recommendation algorithms should be

protected.

A. Why the Court Took This Case

The Court’s decision to hear the case caught many by surprise

(including the authors of this paper). After all, the circuit courts are in

agreement about the application of Section 230 in this context,

dissenting opinions notwithstanding. And, the Supreme Court rarely

grants certiorari to interpret a federal statute in the absence of a circuit

split.

Additionally, the case raises challenging causation issues. The

petitioners argue that platforms are liable for recommending videos

responsible as the ‘develop[er]’ or ‘creat[or]’ of that content.”); see also Dyroff v. Ultimate

Software Grp., 934 F.3d 1093, 1096 (9th Cir. 2019) (“Ultimate Software, as the operator of

Experience Project, is immune from liability under the CDA because its functions, including

recommendations and notifications, were content-neutral tools used to facilitate

communications.”); Marshall’s Locksmith Serv. v. Google, LLC, 925 F.3d 1263, 1270–71 (D.C.

Cir. 2019) (stating that using neutral algorithms—“that do not distinguish between legitimate and

scam locksmiths”—to decide which information appears on a map is protected under Section

230); Carafano v. Metrosplash.com, Inc., 339 F.3d 1119, 1124–25 (9th Cir. 2003) (suggesting

Section 230 protects a platform’s “decision to structure the information provided by users . . . such

as ‘matching’ profiles with similar characteristics”).

36. See infra Part IV.

37. See Question Presented, Gonzalez v. Google LLC, No. 21-1333, 2022 WL 4651229

(2022).

38. See infra notes 40–45 and accompanying text. As we understand it, the term “targeted

recommendations” refers to platforms’ use of recommendation algorithms to personalize content

for their users.

39. See, e.g., Tejas N. Narechania, Certiorari in Important Cases, 122 C

OLUM. L. REV. 923,

927 (2022) (finding it “unusual” for the court to review a case that presented no circuit split).

Admittedly, the Court retains a very broad discretion in its decisions to grant certiorari. Id. at 924.

184 DUKE LAW JOURNAL ONLINE [Vol. 72:176

that a third party posted, which allegedly contributed to inciting the

terrorist attack that led to the death of their relatives. This is a rather

convoluted chain of events, and it is not at all clear that plaintiffs can

successfully establish a causal link between the recommended video

and the attack or the ensuing deaths. More importantly, given the

particular facts of the case, it is questionable whether the Court’s

decision can illuminate the more mundane liability claims that Section

230 regularly shields against, such as garden-variety defamation

actions. Against this background, it is useful to consider why the Court

decided nevertheless to grant certiorari and what kind of changes it

may have in mind.

Reading between the lines, the Court appears to favor limiting

Section 230 immunity for uses of recommendation algorithms. In

granting certiorari in Gonzalez, the Court indicated its willingness to

make broad decisions regarding the application of Section 230 to

recommendation algorithms. The Court framed the question presented

as:

Does section 230(c)(l) immunize [platforms] when they make

targeted recommendations of information provided by third parties,

or only limit the liability of interactive computer services when they

engage in traditional editorial functions (such as deciding whether to

display or withdraw) with regard to such information?

By constructing the question in such an expansive manner, the

Court seems to invite a broad-brush “solution” to the interplay

between Section 230 and recommendation algorithms. This counters a

bottom-up approach that is attentive to different uses and applications

of recommendation algorithms. Taking on Gonzalez to broadly

reshape our understanding of Section 230 is also in line with the

Supreme Court’s recent maximalist tendencies.

Another indication that the Court is inclined to carve out broad

exclusions from Section 230 stems from the general discontent toward

existing regulation of platforms. The Court is no stranger to criticisms

of Section 230 and the power of social media platforms. For example,

Justice Thomas opined on the issue in three previous cases in which the

Court denied certiorari to challenges to Section 230. Against the broad

interpretation of Section 230 that courts have so far adopted, he argued

40. Question Presented, Gonzalez v. Google LLC, No. 21-1333, 2022 WL 4651229 (2022).

41. See generally Strict Scrutiny, This Maximalist Conservative Supermajority, C

ROOKED

MEDIA (June 27, 2022), https://crooked.com/podcast/this-maximalist-conservative-supermajority

[https://perma.cc/4PD5-R9W9] (discussing the Supreme Court’s recent maximalist, rather than

incremental, tendencies in many of the cases it takes on).

2023] GONZALEZ V. GOOGLE 185

that a proper textualist reading would limit Section 230 immunity.

emphasized that upholding immunity protects “unwelcome content”

and that platforms “can solicit thousands of potentially defamatory

statements” while avoiding “product-defect claims” that involve

content about terrorism or human trafficking.

Justice Thomas also

seemed to understand the political economy of Section 230, noting that

a broad reading “confer[s] sweeping immunity on some of the largest

companies in the world.”

Finally, he warned that since private

companies exert “enormous control” over speech, “[w]e will soon have

no choice but to address how our legal doctrines apply to highly

concentrated, privately owned information infrastructure such as

digital platforms.”

B. A Few Warnings

It would be a mistake for the Court to try and “fix” the

shortcomings of online platforms by excluding targeted

recommendations. For one, many (if not all) platforms—including

social media services like Facebook and Twitter and search engines

like Google and Bing—rely upon targeted recommendations to select

and organize content that users will find relevant and engaging.

(And,

of course, to drive advertising revenues.) Using algorithmic tools to

rank and favor content is all but necessary because of the scale and

volume of content uploaded to these platforms.

Indeed, in its early

days Facebook displayed content in reverse chronological order.

But,

as the amount of content it hosted grew gigantically, simple

chronological ordering did not allow users to easily find or process

relevant content. Thus, Facebook started to rely on its newsfeed

algorithm to rank content on users’ behalf, replacing the chronological

ranking with a more sophisticated ranking tools that considers

thousands of relevant factors. Nowadays, Facebook and its users are

42. Doe v. Facebook, Inc., 142 S. Ct. 1087, 1087 (2022) (Thomas, J., concurring);

Malwarebytes v. Enigma Software Grp. USA, 141 S. Ct. 13, 15–17 (2020) (Thomas, J.,

concurring); Biden v. Knight First Amend. Inst. at Columbia Univ., 593 U.S. 1220, 1221 (2021)

(Thomas, J., concurring).

43. Malwarebytes, 141 S. Ct. at 15–18.

44. Id. at 13.

45. Biden, 593 U.S. at 1221.

46. See Rubinstein & Kenneth, supra note 27, at 52–56.

47. Id.

48. A

RAL, supra note 32, at 84.

186 DUKE LAW JOURNAL ONLINE [Vol. 72:176

utterly dependent on algorithmic recommendation to deliver relevant

content to specific users.

Supporting the use of recommendation algorithms is more than a

deferral to companies’ favorite modus operandi. The volume, variety,

and velocity at which online content is generated and processed on

major platforms like Facebook and Google makes it inevitable that

these services rely on recommendation algorithms. It is doubtful that

platforms could provide the benefits that Section 230 hoped to deliver

for users—rich and diverse informational, educational, cultural

resources provided by online speech services—without relying on

recommendation algorithms.

In addition, terminological ambiguity

complicates matters. Platforms have to rely on algorithms to manage

content because of scale, as noted. And, any type of content

management would “recommend” something—be it the

chronologically recent posts or the ones some algorithm deems most

desirable. Hence, as a practical matter, it is not clear what platforms

can do to manage content without using any recommendation

algorithms.

Furthermore, the most common uses of recommendation

algorithms—to favor content that the user is interested in and

connections that the user would like to engage with—are socially

desirable. They make those platforms interesting and engaging for

billions of users with different backgrounds and interests. Put simply,

without recommendation algorithms, large platforms would turn into

ugly assemblages of chaotic, irrelevant, and almost randomly presented

content, depriving users of the value of content recommendations

tailored to their interests.

Note, excluding targeted recommendations from Section 230

immunity will not make the use of recommendation algorithms illegal.

Instead, it would make platforms potentially liable for content that

they recommend. But, this is no small matter. What might happen,

concretely, if the Gonzalez court ends Section 230 immunity for

“targeted recommendations?”

Consider two examples: Facebook and Google. Assume that

Facebook internalizes this regulatory shift and decides to minimize

potential liability by shutting off the recommendation algorithms in its

49. Id.

50. 47 U.S.C. § 230(a)–(b).

51. See infra notes 93–101 and accompanying text; Rubinstein & Kenneth, supra note 27, at

57–61 (arguing recommendation algorithms enable platforms to perform efficient content

moderation and dissemination).

2023] GONZALEZ V. GOOGLE 187

newsfeed algorithm. There are reasons to think that Facebook users

would be worse off. An internal report on the results of an experiment

to this effect found that turning off the newsfeed algorithm “led to a

worse experience almost across the board. People spent more time

scrolling through the News Feed searching for interesting

stuff . . . . They hid 50% more posts, indicating they weren’t thrilled

with what they were seeing.”

Moreover, removing “all ranked

sorting” would probably lead to users seeing even more “borderline”

content than they do with the current system.

As for Google’s search

engine, it too relies heavily on ranked search results.

Indeed, online

search results enjoy expansive legal protection beyond Section 230,

including constitutional safeguards.

However, if Google had to

terminate its use of algorithmic ranking, the quality of its search results

would be diminished beyond recognition.

Of course, this is not to say that recommendation algorithms are

trouble free. Platforms’ uses of recommendation algorithms lead to

many undesirable outcomes. Those include exacerbating body image

problems for teenage girls by promoting images of idealized bodies and

exposing users to undesirable violent and graphic content or

misinformation.

Moreover, online platforms seem very reluctant to

52. See Alex Kantrowiz, Facebook Removed the News Feed Algorithm in an Experiment.

Then It Gave Up, B

IG TECH. (Oct. 25, 2021), https://www.bigtechnology.com/p/facebook-remov

ed-the-news-feed-algorithm [https://perma.cc/AS62-97ZP].

53. See id. (“Wiping out all ranked sorting of the News Feed clearly led to other problems,

including . . . integrity issues.”); Keller, supra note 34, at 256 (pointing out that Facebook’s current

way of handling “borderline” content relies heavily on algorithmic ranking).

54. James Grimmelmann, The Structure of Search Engine Law, 93

IOWA L. REV. 1, 7–11

(2007); see also Danny Sullivan, FAQ: All About the Google RankBrain Algorithm, S

EARCH

ENGINE LAND (June 23, 2016), https://searchengineland.com/faq-all-about-the-new-google-rank

brain-algorithm-234440 [https://perma.cc/2DCA-5BV4] (discussing Google’s use of machine

learning algorithms to help deliver its search results).

55. Courts have recognized First Amendment protections of search results. See, e.g., Best

Carpet Values, Inc. v. Google LLC, No. 5:20-CV-04700-EJD, 2021 WL 4355337 at *10 (N.D. Cal.

Sept. 24, 2021) (citing other cases as well). Compare E

UGENE VOLOKH & DONALD M. FALK,

FIRST AMENDMENT PROTECTION FOR SEARCH ENGINE RESULTS 6–10 (2012) (arguing search

results are entirely protected by the First Amendment), with Oren Bracha & Frank Pasquale,

Federal Search Commission? Access, Fairness, and Accountability in the Law of Search, 93

ORNELL L. REV. 1149, 1193–1201 (2008) (arguing that that the First Amendment does not

encompass search engine results).

56. See Georgia Wells, Jeff Horwitz & Deepa Seetharaman, Facebook Knows Instagram Is

Toxic for Teen Girls, Company Documents Show, W

ALL ST. J. (Sept. 14, 2021) https://www.wsj.

com/articles/facebook-knows-instagram-is-toxic-for-teen-girls-company-documents-show-11631

620739 [https://perma.cc/5BFP-6VRE] (describing how Facebook promoted images of idealized

bodies to teenage girls despite knowing that this exacerbated body image problems for vulnerable

teens); M

OZILLA FOUNDATION, supra note 33 (cataloguing accounts of YouTube’s algorithm

exposing users to undesirable content); see generally Neli Frost, The Global Political Voice Deficit

188 DUKE LAW JOURNAL ONLINE [Vol. 72:176

significantly modify their recommendation algorithms even in light of

such harms.

However, revoking Section 230 immunity for targeted

recommendations is too blunt of an instrument to remedy these failings

and will likely cause other problems. We suggest better approaches in

Part IV.

The previous paragraphs argued against a broad-stroke exclusion

of recommendation algorithms from Section 230 protection. Note,

however, that we make a principled argument against generic

application of Section 230 to recommendation algorithms. As such, we

are also hesitant about adopting a broad-stroke inclusion of

recommendation algorithms under Section 230 protections. Indeed, a

major problem with the Court’s articulation of the question in

Gonzalez is that it seems to force a false dilemma: either regulating

recommendation algorithms by excluding Section 230 or providing

unqualified protections. If required to choose between the two, we

favor retaining Section 230 protections. But, we think this is the wrong

question. As we explain in the next Parts, we favor a more nuanced

analysis—one that extends Section 230 immunities in most cases but

also allows courts to gradually develop exceptions and best practices

that would counter the undesirable effects of Section 230’s private

governance regime.

The bottom line is this: the Court is ill-advised to try and solve the

plethora of problems associated with online speech by rejecting Section

230 immunity for all uses of recommendation algorithms. Such an

outcome would reflect a misunderstanding of relevant technology, a

disregard of the important benefits associated with algorithmic

ranking, and a lack of faith in the ability of legislators and courts (and

platforms) to gradually devise better tailored solutions to the ever-

changing challenges of regulating online speech. This is not only a bad

outcome, it also is clearly misaligned with the purpose of Section 230.

Matrix (Mar. 26, 2022) (unpublished manuscript) (on file with authors) (arguing that amplification

hinders democratic deliberations and other speech-related political interests).

57. See Keach Hagey & Jeff Horwitz, Facebook Tried To Make Its Platform a Healthier

Place. It Got Angrier Instead, W

ALL ST. J. (Sept. 15, 2021), https://www.wsj.com/articles/facebook-

algorithm-change-zuckerberg-11631654215 [https://perma.cc/N8YR-3UYB] (describing how

Facebook’s leadership rejected suggestions to modify its algorithms to deemphasize outrage and

lies because the changes could undermine user engagement); Karen Hao, How Facebook Got

Addicted To Spreading Misinformation, MIT

TECH. REVIEW (Mar. 11, 2021), https://www.technol

ogyreview.com/2021/03/11/1020600/facebook-responsible-ai-misinformation [https://perma.cc/96

6N-2AMB] (describing how Facebook rejected proposals to change its newsfeed algorithm to

reduce political polarization).

58. See infra Part V.

2023] GONZALEZ V. GOOGLE 189

III.

THREE ANALYSES OF RECOMMENDATION ALGORITHMS

In this section we discuss several views that were considered by

lower courts for applying Section 230 to recommendation algorithms

as possible nuanced solutions to Gonzalez v. Google, LLC. We support

one—the majority’s view in the Ninth Circuit—and reject two others.

The analysis that follows culminates in the unsurprising claim that

the Court should uphold immunity for most uses of recommendation

algorithms, including those at issue here. In most respects, we endorse

thirty years of legal reasoning by lower courts about the interpretation

of Section 230. Given the Court’s apparent inclination to make

sweeping changes, we think it is necessary and valuable to highlight

what the status quo gets right and the negative consequences that

would ensue from any radical departures. We recognize that our

preferred solution for Gonzalez does little to address the many perils

of online speech. Later, in the final Part of this paper, we will point to

more adequate ways to address these problems.

A. The Preferred View: Recommendation Algorithms as Tools,

Material Contribution Test

To recap: Section 230 immunizes platforms from being held liable

as publishers or speakers of any information that third-parties publish

and shields platforms from civil liability for voluntarily and in good-

faith restricting access or availability to some materials.

One way to

analyze recommendation algorithms in the context of Section 230

amounts to “business as usual.” On this approach, recommendation

algorithms are “neutral tools” and using them is akin to any other

measure the platforms adopt to manage content. Hence, platforms

would enjoy Section 230 immunity as long as they do not make

“material contributions” to the content that users upload.

Thus

understood, analyzing the use of recommendation algorithms requires

courts to answer two simple questions: are these neutral tools? And,

does the particular use constitute a material contribution to the

content?

We think this approach is the best resolution for Gonzalez. Recall

the plaintiffs in Gonzalez argued that Section 230 does not apply

59. 47 U.S.C. § 230(c); supra Part II.

60. See Fair Hous. Council of San Fernando Valley v. Roommates.com, LLC, 521 F.3d 1157,

1169–71 (9th Cir. 2008) (distinguishing “providing neutral tools” from “materially contributing”

to the alleged unlawfulness); Marshall’s Locksmith Serv. v. Google, LLC, 925 F.3d 1263, 1271

(D.C. Cir. 2019) (holding that algorithms that convert third-party input of location into picture

form use neutral means and therefore enjoy section 230 immunity).

190 DUKE LAW JOURNAL ONLINE [Vol. 72:176

because Google did more than merely publish content. Plaintiffs

argued that the company created and developed the ISIS content that

appears on YouTube.

While the plaintiffs recognized that the

platforms did not initially create the relevant ISIS videos, they still

argued that Google made material contribution by using

recommendation algorithms to match these specific videos to specific

users in order to enhance engagement.

The Ninth Circuit majority

adopted this analysis but rejected the plaintiffs’ claims.

Drawing on

earlier cases, the court found—correctly in our view—that platforms

do not become content creators or developers simply by “supplying

‘neutral tools’ that deliver content in response to user inputs.”

That

is, Google’s recommendation algorithms neither specify nor urge users

to upload any specific content. Rather, as we explained above, these

algorithms analyze users’ behavior on the platform (including posts

and viewing history) and match users with new video recommendations

accordingly in order to enhance engagement.

Hence, the fact that

Google’s (YouTube’s) algorithms recommend ISIS content to users—

based on viewership history, actions, and other information about the

user—should not result in forfeiture of Section 230 immunities.

Similarly, the Second Circuit held in a similar case involving

Facebook’s recommendation algorithms that “[t]he algorithms take

the information provided by Facebook users and ‘match’ it to other

users—again, materially unaltered—based on objective factors

applicable to any content, whether it concerns soccer, Picasso, or

plumbers.”

Whenever platforms use recommendation algorithms only to

“match” between information created by one user to some content

uploaded to the website by other users, platforms should be protected

under Section 230 immunity. Similarly, Section 230 should also protect

adjacent decisions, such as making some content more available than

61. Gonzalez v. Google LLC, 2 F.4th 871, 892 (9th Cir. 2021), cert. granted, No. 21-1333,

2022 WL 4651229 (U.S. Oct. 3, 2022), cert. granted sub nom. Twitter, Inc. v. Taamneh, No. 21-

1496, 2022 WL 4651263 (U.S. Oct. 3, 2022).

62. Id. at 891–93.

63. Id. at 893–97.

64. Id. at 893.

65. Id. at 894–95; see also Carafano v. Metrosplash.com, Inc., 339 F.3d 1119, 1124 (9th Cir.

2003) (“[T]he fact that Matchmaker classifies user characteristics into discrete categories and

collects responses to specific essay questions does not transform Matchmaker into a ‘developer’

of the ‘underlying misinformation.’”). Hence, the website’s decision to match profiles with similar

characteristics is consistent with Section 230 immunity.

66. Gonzalez, 2 F.4th at 894–95.

67. Force v. Facebook, Inc., 934 F.3d 53, 70 (2d Cir. 2019).

2023] GONZALEZ V. GOOGLE 191

others, placing content on specific areas of the website, and deciding

which users will be shown some content based on data about that user.

Those actions should ordinarily be understood as platforms’

management of third-party content for the benefit of the specific users.

They should seldom be regarded as decisions for which platforms

should be held liable.

Using recommendation algorithms to rank content, decide which

should be more visible to (particular) users, and so on are all an

“essential part of traditional publishing.”

These actions do not pass

the line between publishing and speaking and are protected under

Section 230.

As the majority in Force

correctly explained: “Merely

arranging and displaying others’ content to users of Facebook through

such algorithms—even if the content is not actively sought by those

users—is not enough to hold Facebook responsible as the

“develop[er]” or “creat[or]” of that content.”

Even Chief Judge

Katzman, writing an influential partial-dissent in Force, seemed to

agree that in performing these services, Facebook “acts solely as the

publisher.”

On this view, Section 230 immunity extends to using

recommendation algorithms to match content and users, regardless of

the outcomes. Dyroff

is a dire example of this reasoning.

An online

messaging board connected a user, who sought to buy heroin, with

another user who responded to that original message. A day later, the

buyer died because the drugs he bought were laced with fentanyl.

Despite the tragic outcome, the Ninth Circuit held that Section 230

immunities applied.

Just as in the objectionable terrorist content

cases discussed above, the court realized that revoking Section 230

immunity from platforms that use recommendation algorithms is out

of sync with the technology and the law. By using algorithms that

recommend or notify users about information posted on the website,

the Ninth Circuit held in Dyroff that platforms are acting as “publisher

68. Id.

69. Id. at 66–67, 70–71.

70. Force v. Facebook, Inc., 934 F.3d 53 (2d Cir. 2019).

71. Id. at 70.

72. Id. at 82–83, 85 (“Of course, the failure to remove terrorist content, while an important

policy concern, is immunized under § 230 as currently written.”).

73. Dyroff v. Ultimate Software Grp., 934 F.3d 1093 (9th Cir. 2019).

74. Id. at 1093.

75. Id. at 1094–96.

76. Id. at 1097–99.

192 DUKE LAW JOURNAL ONLINE [Vol. 72:176

of other’s content” and should therefore be immune under Section

230.

Lastly, we support extending Section 230 immunity to uses of

recommendation algorithms based on understanding them as “neutral

tools.” For instance, in Force, the Ninth Circuit emphasized that

Facebook’s recommendation algorithms are neutral tools that connect

specific users to specific content and as such are protected under

Section 230.

In our view, when platforms use “content-neutral

algorithms, without more” to match specific content to specific users,

they should retain their Section 230 immunity.

One possible objection to this stance urges that recommendation

algorithms are not really “neutral.” After all, recommendation

algorithms do favor some content: some posts, videos, or groups will

appear at the top of searches or newsfeeds and some at the bottom.

This criticism is untenable. Favoring some content by featuring it more

prominently than other content is unavoidable. Even the Yellow

Pages—which arranges businesses and organizations into groups and

lists their contact information in alphabetical order—makes “Ace

Plumbing” more prominent than “Zeke’s Plumbing.” So, too, for a

platform’s use (and indeed lack of use) of recommendation algorithms.

Any method to manage content would eventually make some content

more prominent. In analyzing those methods, we must look beyond

this feature and evaluate how platforms make those decisions. As

explained, recommendation algorithms are deemed “neutral” because

their curation is not based on the content’s meaning or subject matter.

Rather, they rely on objective factors applicable to any content

(“whether it concerns soccer, Picasso, or plumbers”

) to decide which

content to amplify.

For those reasons, we think recommendation algorithms are

usually content-neutral for First Amendment purposes.

And, for

similar reasons, courts have correctly recognized that using content

77. Id. at 1098 (“These functions—recommendations and notifications—are tools meant to

facilitate the communication and content of others.”).

78. Force, 934 F.3d at 66–67, 70–71.

79. Gonzalez v. Google LLC, 2 F.4th 871, 897 (9th Cir. 2021), cert. granted, No. 21-1333,

2022 WL 4651229 (U.S. Oct. 3, 2022), cert. granted sub nom. Twitter, Inc. v. Taamneh, No. 21-

1496, 2022 WL 4651263 (U.S. Oct. 3, 2022); Force, 934 F.3d at 69–70; Marshall’s Locksmith Serv.

v. Google, LLC, 925 F.3d 1263, 1270–71 (D.C. Cir. 2019); Fed. Trade Comm’n v. LeadClick

Media, LLC, 838 F.3d 158, 174–79 (2d Cir. 2016).

80. Force, 934 F.3d at 70.

81. Id.

82. Rubinstein & Kenneth, supra note 27, at 56.

2023] GONZALEZ V. GOOGLE 193

recommendation algorithms for such purposes does not forfeit Section

230 immunity.

* * *

As noted, influential dissenting opinions have rejected the

“business as usual” approach.

Objecting to the extensive protection

that Section 230 provides to platforms under the material contribution

standard, they sought to limit such protections. Wisely, they realized

that it makes little sense to limit those protections by excluding

recommendation algorithms from Section 230 protections altogether.

Instead, they tried to single out specific, yet still too broad, applications

of recommendation algorithms by platforms and explain why those

should be excluded from Section 230 immunity. We disagree with both

dissenting views on the merits, as we explain below.

B. First Alternative View: Recommending Connections

One view favors excluding recommending “connections” from

Section 230 protections. Judge Berzon, writing concurrently in

Gonzalez, argued that platforms forfeit their Section 230 immunity

when they amplify and direct content to specific users. That they use

“neutral” algorithms to do so matters little in her view.

“These types

of targeted recommendations and affirmative promotion of

connections and interactions among otherwise independent users,” she

opined, “are well outside the scope of traditional publication.”

According to this view, when recommendation algorithms are used to

facilitate connections and social interactions, they are not protected by

Section 230. Is there a persuasive explanation for this conclusion? We

think not.

One possible explanation is that connections on platforms are a

kind of content, an input that users upload to the platform. On this

view, recommending connections to users implies participating in the

creation of the content. That is, when recommendation algorithms are

applied to connections, they always amount to a material contribution

and thus are never protected under Section 230. On this account, when

users connect to users or groups, they upload content implicitly stating,

“I like this group/user and want to connect with them.” In turn, when

83. See, e.g., Force, 934 F.3d at 76 (Katzmann, C.J., concurring); Gonzalez, 2 F.4th at 918

(Gould, J., concurring).

84. Gonzalez, 2 F.4th at 914.

85. Id.

194 DUKE LAW JOURNAL ONLINE [Vol. 72:176

platforms use recommendation algorithms to suggest specific friends,

groups, or events, they implicitly tell users, “I think you will like X,”

and the user implicitly responds, “I’m following your advice; I do like

X.”

There are several problems with this account. To begin with, we

are hesitant to say that connections are themselves content. While they

are created by users’ input, connections seem to be more part of the

structure of the platform than something that users try to convey to

others. Moreover, this argument seems to suggest that

recommendation algorithms are also content. Allegedly, the use of

recommendation algorithms converts the content of “connection to X”

to “platform thinks you will like connection to X.” But, ascribing such

content to recommendation algorithms is mistaken. Such uses of

recommendation algorithms only help platforms decide which

connections they should recommend to which user. As the Ninth

Circuit held in Dyrrof, “[T]hese functions—recommendations and

notifications—are tools meant to facilitate the communication and

content of others. They are not content in and of themselves.”

Thus,

we do not think that recommending content should be understood as

contributing to creation of content.

Even if recommending connections is somehow contributing to

content, that would not suffice. To decide whether Section 230

protections apply, we must find that what platforms do with those

recommendations amounts to a material contribution. We are hesitant

to agree that recommending connections always amounts to material

contribution.

There are many ways to implement connection recommendations

on the platform, some more pervasive than others. Without

considering more details about the means platforms use to recommend

connections to specific users—how often do these recommendations

appear, how much screen space do they capture, how easy it is for users

to ignore, how often do users actually ignore those recommendations,

etc.—it is difficult to say whether these recommendations amount to a

material contribution.

We can imagine that some uses of

recommendation algorithms can make material contributions. Most

86. Dyroff v. Ultimate Software Grp., 934 F.3d 1093, 1098 (9th Cir. 2019) (holding that

Section 230 immunized social networking operator from liability for its alleged role in facilitating

the drug overdose death of a man who used the social network to identify a local drug dealer and

obtain heroin, which turned out to be laced with fentanyl).

87. Compare Fair Hous. Council of San Fernando Valley v. Roommates.com, LLC, 521 F.3d

1157, 1169–71 (9th Cir. 2008), with Marshall’s Locksmith Serv. v. Google, LLC, 925 F.3d 1263,

1271 (D.C. Cir. 2019); Dyroff, 934 F.3d at 14–17.

2023] GONZALEZ V. GOOGLE 195

obviously, if a platform requires a user to connect to one group or

friend out of the recommended list in order to continue using the

platform, it would likely amount to material contribution excludable

from 230 immunity.

But, for similar reasons, we are doubtful that

merely using recommendation algorithms to recommend connections

on platforms, without additional information, suffices to forfeit Section

230 protections. As we argue throughout this Article, a more nuanced

analysis of the application of recommendation algorithms is necessary.

Finally, we do not argue that by using recommendation algorithms

platforms are conveying a message. We suggest that platforms would

lose protection under Section 230 when they materially contribute to

content that a user uploads. This view accepts the existing analysis of

Section 230, which focuses on the material contribution to the content

that users upload. Using recommendation algorithms to suggest

connections can sometimes help create or develop this content.

Whether such a contribution is sufficient to strip platforms of their 230

immunity, though, requires further details and a case-by-case

approach.

C. Second Alternative View: Conveying a Message

Another view of recommendation algorithms detaches them from

users’ content entirely. In Force, Chief Judge Katzman seemed to

support this view. He argued that recommendation algorithms that

match different users with similar interests do more than “publish”

users’ content.

Rather, they “forge[] connections, [and] develop[] new

social networks.”

In his words, when “targeting and recommending

[profile, group, or event pages written by other users] to users,”

Facebook “uses the algorithms to create and communicate its own

message.”

So, the argument goes, using recommendation algorithms

to suggest friends and groups is not protected under Section 230

because these activities amount to conveying messages, not merely

publishing them.

88. Cf. Roommates.com, 521 F.3d at 1166 (“By requiring subscribers to provide the

information as a condition of accessing its service, and by providing a limited set of pre-populated

answers, Roommate becomes . . . the developer, at least in part, of that information.”). Because

Roommates.com made material contributions to the content, it was not protected under Section

230. Id.

89. Force v. Facebook, Inc., 934 F.3d 53, 76–77 (2d Cir. 2019).

90. Id.

91. Id. at 82.

92. Id. at 76–77, 82.

196 DUKE LAW JOURNAL ONLINE [Vol. 72:176

Following this analysis, the Supreme Court might find that

recommendation algorithms are not neutral tools but rather are tools

that actively and deliberately convey information. Drawing on the

colloquial language of “recommendation algorithms,” the Court might

say that these algorithms are in fact just that: a recommendation, a

message that platforms convey to users about some content. Granted,

if using recommendation algorithms conveys a message, then it is not

protected under Section 230. But, is this interpretation compelling? Is

the use of recommendation algorithms really conveying a message? We

are doubtful.

As we explained elsewhere, even if platforms have messages that

they wish to convey, it is unclear how the use of recommendation

algorithms to rank content gives a voice to these messages.

It is tempting to understand ranking and moderating content as

complementary activities since both involve the selection,

organization, and presentation of online content, or what many refer

to as curation. And, if ranking is a form of content curation, then it also

seems to involve the exercise of editorial discretion (“we recommend

this, not that”) and, therefore, convey a message. This reading suggests

that platforms should be treated as publishers (because both make

editorial decisions) and, thus, possibly liable for information uploaded

to their websites, despite Section 230(c)(1).

But, the dissimilarities between algorithmic ranking by platforms

and the editorial decisions of traditional media outlets (like

newspapers) are striking. Editors are responsible for the content and

style of a newspaper. They assign, review, edit, rewrite, and lay out all

copy, drawing on their communication and writing skills, their

familiarity with various issues, policies, and events, and their subject-

matter expertise while maintaining their independence. In designing

ranking algorithms, however, technical teams engage in none of these

tasks. Rather, they use complex mathematics and sophisticated

engineering techniques to determine in a computationally efficient

manner which of many personal characteristics are most relevant for

predicting engagement with available content.

Thus, algorithmic recommendation, unlike content-moderation,

does not entail policy formation, decisions about values and

viewpoints, or human oversight of automated judgments to ensure

fidelity with these editorial standards. For example, Facebook’s

recommendation algorithms evaluate thousands of pieces of content

93. See Rubinstein & Kenneth, supra note 27, at 57–61.

2023] GONZALEZ V. GOOGLE 197

based on hundreds of signals to determine which content is most likely

to keep each of the platform’s billions of users most engaged. It is

doubtful that Facebook’s values and viewpoints are a good predictor

of what a particular user might find relevant. In a nutshell, why would

Facebook’s (or Nick Clegg’s) core values matter if we are trying to

predict whether a particular user is more engaged by dog photos or cat

photos or by John Stuart Mill or Edmund Burke?

Hence, even if

platforms wanted to instill their values as part of the algorithmic

recommendation process, those would probably have little weight in

light of the algorithm’s major task: keeping the user engaged.

Furthermore, given the sheer scale of the platforms in question

and the number and variety of possible topics they recommend to users

at any given time, it is doubtful that we can intelligibly identify “the

message” that platforms convey, let alone ascribe platforms any

meaningful intention to convey it.

The only way to derive a message

from all this recommended content would be to analyze all of it at a

very high level of generality and abstraction, which arguably brings it

back within the ambit of Section 230 immunity.

It follows that any attempt to extract a coherent message out of

the varied content that platforms recommend to billions of different

users is a fool’s errand. At most, one can say that recommendation

algorithms endorse a specific user’s engagement with some specific

content. But, this is a very limited form of endorsement, one that

cannot easily be traced to platform’s attempt to convey a message

about any of the topics it recommends. Rather than conveying “the

platform’s message,” the underlying algorithms are designed to

amplify whatever content is likely to engage the user.

In Gonzalez, Judge Gould subscribed to the view that

recommendation algorithms convey a message but concluded that

platforms should only forfeit Section 230 immunity when the

94. Id. Clegg is Meta’s President of Global Affairs. Nick Clegg, President, Global Affairs,

ETA, https://about.meta.com/media-gallery/executives/nick-clegg [https://perma.cc/C8J7-FKZ7].

95. Indeed, the platform’s values and the user’s interest might be connected. Arguably,

many people opt for Facebook over Gab (for instance) exactly because the former provides

content that they are interested in watching, while the latter provides content that they do not

want to see. So, the platform’s “values” and the user’s explicit interests (which might be what

actually keeps a user engaged) may be confounding factors. However, even under this account,

the recommendation algorithm recommends some content because it keeps the user engaged, not

because it serves the platform’s values.

96. Admittedly, a platform can convey a message, like when it overrules the ordinary uses

of its algorithms and inputs a message such as “go vote” or “get vaccinated.” However, these are

not the actions of recommendation algorithms. If anything, they show the platform overriding

them.

198 DUKE LAW JOURNAL ONLINE [Vol. 72:176

information they amplify is particularly problematic.

With this in

mind, he argued that recommendation algorithms should not be

protected “because of the unique threat posed by terrorism

compounded by social media.”

He adopted an explicitly content-

based approach, holding that courts should be able to hem-in

recommendation algorithms that amplify bad messages.

But, this

approach runs afoul of the ideas underpinning Section 230.

100

Holding

that Section 230 immunity applies unless platforms amplify content

that is “very bad” provides platforms very little assurances regarding

what they can and cannot publish online without risking liability. The

whole point of legislating and interpreting Section 230 broadly was to

avoid this uncertainty and the expected chilling effects that would

likely follow.

101

As we shortly explain, we are not necessarily opposed

to carving out specific topics from Section 230 protection, but this

approach must be limited and narrowly framed.

IV.

CONCLUDING REMARKS: BETTER WAYS TO REGULATE ONLINE

SPEECH

The foregoing discussion offers several valuable lessons for the

relation between Section 230 and recommendation algorithms. First

and foremost, recommendation algorithms are best understood as a

method that platforms apply for various uses. As such, courts (and

legislators) should refrain from regulating all recommendation

algorithms generically. Instead, courts should opt for a more nuanced

97. Gonzalez v. Google LLC, 2 F.4th 871, 920–21 (9th Cir. 2021), cert. granted, No. 21-1333,

2022 WL 4651229 (U.S. Oct. 3, 2022), cert. granted sub nom. Twitter, Inc. v. Taamneh, No. 21-

1496, 2022 WL 4651263 (U.S. Oct. 3, 2022).

98. Id. at 923 (“I would hold that where the website (1) knowingly amplifies a message

designed to recruit individuals for a criminal purpose, and (2) the dissemination of that message

materially contributes to a centralized cause giving rise to a probability of grave harm, then the

tools can no longer be considered ‘neutral.’”).

99. Id. at 921 (“[T]he seemingly neutral algorithm instead operates as a force to intensify

and magnify a message . . . . But when it shows acts of the most brutal terrorism

imaginable . . . then the benign aspects of Google/YouTube, Facebook and Twitter have been

transformed into a chillingly effective propaganda device . . . .”).

100. As the court stated,

But this is not where Congress drew the line . . . . Congress did not differentiate

dangerous, criminal, or obscene content from innocuous content when it drafted §

230(c)(1). Instead, it broadly mandated that ‘[n]o provider . . . of an interactive

computer service shall be treated as the publisher or speaker of any information

provided by another information content provider.’

Id. at 896 (quoting 47 U.S.C. § 230(c)(1)); see also id. at 912; supra Part I.

101. See Goldman, supra note 10, at 155–57 (noting Congress sought to incentivize platforms

to moderate objectionable content within Section 230).

2023] GONZALEZ V. GOOGLE 199

approach, one that considers the specific use and application of

recommendation algorithms in specific contexts. This argument is far

from obvious. Against the background of growing discontent over the

monopolistic power of platforms and the ways they manage content,

many think that Section 230 should be scaled back.

102

This might be the

case. But, excluding recommendation algorithms from Section 230

immunity is not the way to go.

Second, lower courts’ focus on the material contribution standard

seems justified. When platforms do not interfere with users’ discretion

to decide which content to upload, platforms should be regarded as

publishers and enjoy Section 230 immunity.

103

Conversely, when

platforms drive users to upload specific content, the platforms make a

material contribution to the content and thus lose such immunity. This

approach is harmonious with the broad text and interpretation of

Section 230. And, it allows courts to analyze the specific application of

recommendation algorithms—whether they were indeed used as

neutral tools in a specific context. This was also the dividing line

between cases like Dyroff and Carafano

104

on one hand and

Roommates.com

105

on the other. Obviously, this legal standard is not

perfect, and the line between material contribution and editorial

judgement is blurry. But, these are ordinary questions of interpretation

that courts regularly address.

106

Third, courts should reject views that treat recommendation

algorithms as inherently conveying a message. As explained, this view

misconstrues the technology behind the recommendation algorithms

and misunderstands the scale and volume at which platforms manage

content nowadays.

Finally, we recognize that our approach does not resolve the many

problems that recommendation algorithms and content amplification

pose in online platforms.

107

We do not think it should. Gonzalez is not

102. See supra notes 17–19 and accompanying text.

103. See, e.g., Fair Hous. Council of San Fernando Valley v. Roommates.com, LLC, 521 F.3d

1157, 1166 (9th Cir. 2008); Force v. Facebook, Inc., 934 F.3d 53, 67 (2d Cir. 2019); Gonzalez, 2

F.4th at 892–93.

104. Carafano v. Metrosplash.com., Inc., 339 F.3d 1119 (9th Cir. 2003).

105. Fair Hous. Council of San Fernando Valley v. Roommates.com, LLC, 521 F.3d 1157 (9th

Cir. 2008).

106. Arguably, the platforms always influence the kind of content that is uploaded—using

content moderation, banning specific words, cultivating specific culture, etc. Additionally, content

moderation schemes that impose various sanctions on specific content influence users to upload

only complying content ex ante. The question, therefore, is to what extent and how blunt is the

platforms’ intervention. This is a challenging question, but one that courts are used to facing.

107. See supra Part II.A.

200 DUKE LAW JOURNAL ONLINE [Vol. 72:176

the case to “solve” Section 230. If the Supreme Court in Gonzalez

decides that recommendation algorithms are immune under Section

230 in most circumstances, there are still viable—indeed desirable—

solutions to the perils of online speech. Discussing each of those would

require separate papers, but we present below a few options we find

appealing.

For one, nothing prevents Congress (politics aside) or the Court

from imposing new restrictions on the use of social media by ISIS and

other terrorist organizations. We are not opposed to carving out

exceptions to Section 230 immunity. But, we think those exceptions (to

the extent that they are desirable on the merit) should be grounded in

and limited to very clearly defined categories. This approach is

exemplified in Section 230’s exception for sex trafficking, which refers

to civil and criminal offenses under designated statutes.

108

For instance,

given the circumstances of Gonzalez, it seems reasonable to fashion a

similar narrow exception to Section 230 that would forfeit protections

to platforms for failing to adequately confront terrorism as defined

under the ATA.

109

Moreover, we think that a carefully written law that

narrowly regulates the use of recommendation algorithms in those

specific contexts might survive First Amendment scrutiny.

110

However,

a judicial decision stripping Section 230 protections for any use of

targeted recommendations (as the Supreme Court seems to

contemplate) or one limiting those protections when “harmful

content” is at play (along the line of Judge Gould’s view) is simply too

broad and should be avoided.

Another viable alternative is to amend Section 230 by

conditioning immunity on the platform demonstrating that it has taken

“reasonable steps to prevent or address” unlawful uses of its services.

This approach permits the courts to decide whether the steps taken by

a service in a given case were reasonable or negligent under the

circumstances in question.

111

In turn, this approach invites the kind of

nuanced analysis of the methods that a particular platform used with

108. See 47 U.S.C § 230(e)(5).

109. Some legislative proposals have taken this approach. See, e.g., Protecting Americans

from Dangerous Algorithms Act, H.R. 2154, 117th Cong. (2021) (removing Section 230 immunity

from large social media companies that amplify or recommend content that is directly relevant to

a claim involving civil rights or acts of international terrorism under the ATA). Note that there

are many problems with the definition of “terrorism,” making it perhaps too flexible of an

exception to Section 230.

110. See Rubinstein & Kenneth, supra note 27, at 61–62 (arguing that targeted regulation of

platform amplification mechanisms that pursue compelling government interests could survive

First Amendment scrutiny).

111. See Citron & Wittes, supra note 9, at 419.

2023] GONZALEZ V. GOOGLE 201

regard to the particular content. As discussed, we think this nuanced

approach is desirable. Yet another option is to adopt a model that

European and other countries have embraced: soft-law mechanisms

that influence platforms to self-regulate and enforce their policies in a

manner conducive to the specific relevant harms.

112

Also, we think that

both platforms and regulators should explore the use of innovative

solutions, specifically ones that challenge the engagement-based

business model and technological architecture of online platforms. In

this sense, introducing “friction”

113

or “middleware”

114

to the online

platforms landscape seems worthwhile.

ONCLUSION

In closing, the analysis of recommendation algorithms in this

Article only applies to Section 230. Courts should be cautious and

deliberate about their use of the preceding analysis in other contexts,

such as First Amendment law. We have argued elsewhere that the

regulation of recommendation algorithms by Florida’s social media law

and certain proposed federal legislation is content-neutral for First

Amendment purposes.

115

But, much depends on the wording and

precise motivation of these provisions. As always, the devil is in the

details.

112. See Rubinstein & Kenneth, supra note 27, at 35–49 (discussing the use of soft-law

measures to confront online public health misinformation).

113. See, e.g., E

RIN SIMPSON & ADAM CONNER, CTR. FOR AM. PROGRESS, FIGHTING

CORONAVIRUS MISINFORMATION AND DISINFORMATION: PREVENTIVE PRODUCT

RECOMMENDATIONS FOR SOCIAL MEDIA PLATFORMS 10 (2020) (recommending that platforms

voluntarily adopt friction measures to hinder amplification of public health misinformation).

114. See, e.g., Francis Fukuyama, Making the Internet Safe for Democracy, 32 J.

DEMOCRACY

37, 40 (2021) (outlining a proposal “to outsource content curation from the dominant platforms

to a competitive layer of ‘middleware companies’”).

115. See Rubinstein & Kenneth, supra note 27, at 51–52, 56–61 (analyzing amplification or

ranking of social media posts as a content-neutral task); see also NetChoice, LLC v. Att’y Gen.,

34 F.4th 1196, 1226 (11th Cir. 2022) (noting that a provision allowing users to opt-out of platform

recommendations of content “is pretty obviously content-neutral”).