Introduction

When an individual first visits video sharing platform TikTok, TikTok starts playing a stream of videos that may seem random without any input from the individual. Over time, however, TikTok will "learn" a viewer's preferences and start showing more specific videos tailored to those preferences.

Journalists have written extensively about TikTok's "For You" landing page and the process by which the platform hones its video recommendations. Frequently, this process is described in lofty language, as in an article from the New York Times titled "How TikTok Reads Your Mind."¹ An investigation from the Wall Street Journal, which involved creating 100 automated accounts programmed with certain interests, reported that some of its accounts "ended up lost in rabbit holes of similar content, including one that just watched videos about depression."² For all the mystery and bravado used to characterize TikTok's For You page, computer scientists have described TikTok's systems as "pretty normal."³ The use of algorithms to recommend content is widespread on the internet, including outside social media.

In recent years, Congress has paid special attention to the interaction between algorithms and content hosted and displayed by online services. One topic of recurring interest is the role online services play in "promoting" or "amplifying" content, particularly content that may be unlawful or harmful. Congress has focused on TikTok's For You page specifically,⁴ but has also expressed concern with algorithm-based recommendation systems more generally.⁵

A question relating to the amplification of harmful material is whether an online service may face legal liability for using algorithms to recommend content to the service's users. The victims and families of victims of terrorist attacks have filed several lawsuits against social media providers alleging that the providers contributed to terrorism by recommending terrorist content. None of these lawsuits has been successful. One obstacle is Section 230 of the Communications Act,⁶ a federal law that provides legal immunity for providers of "interactive computer services" (ICS) in certain circumstances. Social media providers have successfully argued that algorithm-based recommendations are protected from liability under Section 230.

This report provides an overview of the current landscape of legal liability for algorithm-based recommendations. It begins with a discussion of algorithms and recommendation systems, followed by a discussion of Section 230 and relevant caselaw involving recommendations by algorithm. The report concludes with a discussion of considerations for Congress, including implications of recent Supreme Court decisions involving Section 230, possible questions related to legislative proposals, and potential issues raised by the Free Speech Clause of the First Amendment.

Background

As the term is commonly used in relation to online services, an algorithm is a problem-solving process undertaken by a computer.⁷ A variety of automated tasks are accomplished via algorithms. These tasks can include producing expressive outputs (such as algorithms that determine a response provided by a virtual assistant or chatbot)⁸ as well as performing mechanical tasks (such as an algorithm that determines when a car's antilock braking system will trigger). As discussed below, how an algorithm operates, and the outputs it generates, may impact whether certain legal protections are available for the algorithm's operator.⁹

In 2023, the internet features a wide variety of online platforms¹⁰ that offer content to the platforms' users. The content available on these platforms may be either first-party content created by the platform or third-party content created by someone other than the platform itself, including user-generated content created by users of the platform. Some platforms may host both first-party and user-generated content, such as an online publication that allows users to leave comments on articles. Other platforms, such as online marketplaces and video streaming services, may host both first-party content and third-party content not provided by users (such as product listings by third-party sellers or video programming provided by third-party licensees).¹¹

Many platforms use algorithms to organize and recommend content. One popular example that has persisted since the early days of the internet is the use of algorithms by search engines to provide the most relevant results to a user's search query.¹² Different search engines may give different weights to a variety of factors in ordering their results, including user-specific information like geographic location and past searches as well as general information like how a search is worded and the popularity of particular websites in response to similar searches.¹³ In addition to search engines, which use algorithms to filter third-party content, platforms that host both first-party and third-party content may use algorithms to sort their content. For example, online marketplaces may use algorithms to determine how products are displayed, and video streaming services may use algorithms to determine what programming to suggest to users.

Platforms that primarily distribute user-generated content, such as social media platforms, may use algorithms to determine what content is displayed to a particular user.¹⁴ These algorithms can support various forms of content distribution, ranging from allowing users to sort content by how often it has been viewed to targeted online advertisement relying on a wide range of factors. Social media platforms may rely on more complicated algorithms in combination with the use of specific user-selected criteria to determine how content is displayed. For example, social media platform Instagram shows its users a "Feed" that prioritizes content from individuals "followed" by the user.¹⁵ Some platforms may place less weight on user-selected criteria: for example, TikTok's For You page displays content regardless of whether the user has followed any specific accounts.¹⁶

Social media platforms are seen as operating along a spectrum of algorithmic curation, with some platforms relying wholly or mostly on more complex processes to determine how content is displayed and others relying on user-defined inputs such as "follows" or user-selected topics.¹⁷ Computer science experts use the term "recommendation system" or "recommender system" to refer to these collections of algorithms.¹⁸ Some commentators have used the term "algorithmic amplification," referring to the use of algorithms to increase the prominence of particular content.¹⁹

Statutory Background: Section 230

Courts have largely rejected attempts to sue platforms for algorithms that organize and recommend user content. Section 230 of the Communications Act of 1934, enacted as part of the Communications Decency Act of 1996, provides legal immunity to providers and users of interactive computer services (ICS), a defined term discussed below.²⁰ Two provisions of Section 230 provide the primary framework for this immunity. The first of these provisions, 47 U.S.C. § 230(c)(1), specifies that ICS providers and users may not "be treated as the publisher or speaker of any information provided by another information content provider." The second provision, Section 230(c)(2), states that ICS providers and users may not "be held liable" for voluntary, "good faith" actions "to restrict access to or availability of material that the provider or user considers to be obscene, lewd, lascivious, filthy, excessively violent, harassing, or otherwise objectionable, whether or not such material is constitutionally protected."

Thus, Section 230(c)(2) more narrowly focuses on actions restricting certain types of objectionable content, while Section 230(c)(1) provides broader immunity for acting as a "publisher or speaker" of another's content.²¹ Defendants invoke Section 230(c)(1)'s broader immunity much more frequently, particularly because a number of courts have interpreted Section 230(c)(1) to also encompass actions to remove or restrict content.²² Accordingly, although an algorithm may be used both to promote and restrict access to content, most cases considering whether Section 230 protects the use of recommendation algorithms have focused on Section 230(c)(1).

A court's decision to apply Section 230(c)(1) to bar legal liability depends on the presence of three conditions.²³ First, the defendant must be a user or provider of an interactive computer service. Second, the liability must arise from the defendant acting as a publisher or speaker. Third, the liability must arise from information provided by another person.

Courts around the country have written decisions in Section 230 cases addressing how these conditions might be satisfied. Without Supreme Court precedent interpreting Section 230, federal and state courts frequently rely on interpretations and analyses undertaken by federal appellate courts.²⁴

Interactive Computer Service

Section 230 defines an ICS as "any information service, system, or access software provider that provides or enables computer access by multiple users to a computer server."²⁵ This definition is broad and applies to more than just websites or social media platforms. Courts have considered various online platforms such as Craigslist,²⁶ Facebook,²⁷ GoDaddy,²⁸ Yahoo!,²⁹ and Zoom³⁰ to be ICS providers.³¹ They have also held that providers of broadband services (e.g., AT&T, Verizon)³² and search engines (e.g., Google)³³ qualify as ICS providers.³⁴ Because courts have construed the definition of interactive computer service broadly, the success of a Section 230(c)(1) defense more often turns on whether the plaintiff seeks to hold the defendant liable as a publisher or speaker and whether the plaintiff's claim arises from information provided by another information content provider.

Role as Publisher or Speaker

Section 230(c)(1) prohibits courts from treating an ICS provider as a "publisher or speaker" of third-party content.³⁵ Courts have interpreted this provision to apply broadly to bar any claim arising from third-party content. Courts have declined to apply Section 230(c)(1) to claims that rely on a provider's own unlawful conduct, rather than its publication of third-party information.

Treatment As "Publisher or Speaker"

In determining whether a legal claim treats a service provider as a "publisher or speaker," courts often look to Zeran v. America Online, an early Fourth Circuit case applying Section 230.³⁶ The Zeran court determined that Section 230(c)(1) bars "lawsuits seeking to hold a service provider liable for its exercise of a publisher's traditional editorial functions—such as deciding whether to publish, withdraw, postpone, or alter content."³⁷ In reaching this conclusion, the Zeran court determined that use of the term "publisher" in Section 230(c)(1) was meant to extend protection to all entities engaged in such functions.³⁸ In Zeran, the Fourth Circuit held that an ICS provider's hosting of defamatory messages and failure to remove them upon receiving notice of their allegedly defamatory nature was protected activity under Section 230(c)(1).³⁹ Many courts have used the Zeran court's description of "traditional editorial functions"⁴⁰ to determine whether a claim would impermissibly treat a service provider or user as a publisher or speaker of another's content.⁴¹

Although courts have continued to rely on Zeran, the breadth of this description of "publisher" activity has concerned some jurists. A number of judges have questioned whether Section 230(c)(1)'s application has expanded beyond its intended scope—although few of these judges have altered the prevailing legal standard.⁴² In a statement respecting a denial of certiorari in a Section 230 case, Justice Clarence Thomas suggested that Zeran and other cases employing its traditional editorial functions analysis had "extend[ed] § 230 beyond the natural reading of the text . . . ."⁴³ Justice Thomas's analysis relied among other things on the use of the terms "publisher" and "distributor" in Stratton Oakmont, Inc. v. Prodigy Services Co., a pre-Section 230 case that at least in part provided the impetus for Section 230's passage.⁴⁴ A defamation case, Stratton Oakmont distinguished between "publishers" liable for defamatory statements and "distributors" liable only if they know or have reason to know of the defamatory statements.⁴⁵ Zeran explicitly rejected the distinction between publishers and distributors and held that Section 230(c)(1)'s protection encompasses "distributor" activity.⁴⁶

Apart from these individual judges, one recent federal appeals court opinion appeared to narrow Zeran's conception of "publisher" activity. In Henderson v. Source for Public Data, the Fourth Circuit held that to treat a service provider as a publisher or speaker, a claim must hold a service provider liable based on the improper content of the disseminated information.⁴⁷ The court drew this requirement from defamation law, under which a defendant's liability as a publisher depends on the improper, "false and defamatory" nature of the material published.⁴⁸ Under this view, Section 230 did not bar claims alleging that a website had failed to comply with the Fair Credit Reporting Act.⁴⁹ Although the claims would have held the site liable for improperly disseminating information, they did not depend on the information's content being improper.⁵⁰ The opinion cited and purported to apply Zeran, but Henderson appeared to add a new requirement since Zeran made no reference to the content of information.⁵¹ Several decisions from state and federal courts outside of the Fourth Circuit have declined to follow Henderson, observing that the decision conflicts with binding precedent in their jurisdiction that reads Section 230(c)(1) more broadly.⁵²

Non-Publisher Activity

Section 230(c)(1) does not bar claims arising from a platform's own non-publisher conduct, though courts have disagreed over the exact boundaries between actionable non-publisher conduct and protected publisher activity. In one case, the Ninth Circuit determined that claims brought against the maker of Snapchat for negligently designing its platform to include a "speed filter" that encouraged users to drive at recklessly high speeds would not be barred by Section 230(c)(1).⁵³ The Ninth Circuit determined that the claims based on Snapchat's speed filter did not treat the platform as a "publisher or speaker," because the claims "treat[ed] Snap as a products manufacturer, accusing it of negligently designing a product (Snapchat) with a defect . . . ."⁵⁴ A state court in Georgia reached a similar conclusion, holding that claims based on Snapchat's speed filter "do not seek to hold Snapchat liable for publishing" and therefore could proceed.⁵⁵

When a product feature determines how user content is displayed or sorted, courts are more likely to determine that a claim treats a service provider as a "publisher or speaker" entitled to protection under Section 230(c)(1).⁵⁶ In an early case on the topic, the Fifth Circuit affirmed the dismissal of a lawsuit alleging that MySpace acted negligently in failing "to implement basic safety measures to prevent sexual predators from communicating with minors on its Web site."⁵⁷ The court concluded that the plaintiff's allegations were "merely another way of claiming that MySpace was liable for publishing [predators'] communications."⁵⁸ In the court's view, the negligence claims hinged on MySpace's publisher functions: its decisions relating to the "monitoring, screening, and deletion" of third-party content.⁵⁹

Information Provided by Another Information Content Provider

The third criteria for Section 230(c)(1) immunity is that liability arises from information "provided by another information content provider."⁶⁰ Put another way, a user or provider of an interactive computer service cannot claim Section 230(c)(1)'s protection for its own content.⁶¹ Under Section 230, a user or provider is considered an "information content provider" of particular content if the user or provider is "responsible, in whole or in part, for the creation or development" of the content.⁶² A defendant cannot rely on Section 230(c)(1) for claims based on content it has created or developed, as such content is not provided by "another" information content provider. Notably, a service provider or user can be merely a publisher of another's information in some circumstances but a content provider in others.⁶³ Whether Section 230(c)(1) applies depends on the particular content being challenged.⁶⁴

"Material Contribution" and "Neutral Tools"

A foundational case on whether a service provider is responsible for particular content is the Ninth Circuit's decision in Fair Housing Council of San Fernando Valley v. Roommates.com, LLC (Roommates).⁶⁵ That opinion said "a website helps to develop unlawful content [and is therefore unable to claim protection under Section 230(c)(1)] if it contributes materially to the alleged illegality of the conduct."⁶⁶ In a later Ninth Circuit opinion, the court clarified that this "material contribution" test "draw[s] the line at 'the crucial distinction between, on the one hand, taking actions (traditional to publishers) that are necessary to the display of unwelcome and actionable content and, on the other hand, responsibility for what makes the displayed content illegal or actionable.'"⁶⁷ Even the act of publishing itself may materially contribute to the unlawfulness of conduct if the content published is private information legally protected from disclosure.⁶⁸

In another portion of the Ninth Circuit's Roommates decision, the court opined that "passive conduits" or "neutral tools," such as a search engine that filters content only by user-generated criteria, would not be responsible for developing content where they do not enhance the unlawfulness of the content.⁶⁹ By contrast, for example, the court said that where a website edited user-generated content in a way that made the message libelous, the site would be "directly involved in the alleged illegality and thus not immune."⁷⁰ In one application of this "neutral tools" analysis, the Ninth Circuit held that customer review aggregator Yelp's rating system, which transforms aggregated user input into a 0-5 star rating, did not amount to development.⁷¹ In another case, the D.C. Circuit determined that social media platform Facebook "provides a neutral means by which third parties can post information of their own independent choosing online," and Facebook was therefore not responsible for developing third-party content merely for failing to remove it.⁷²

Section 230 Immunity and Algorithmic Recommendations

Individuals have occasionally sought to hold social media platforms and search engines—both of which federal courts have held to be providers of "interactive computer services" under Section 230⁷³—liable for their use of algorithms or other automated systems to recommend or organize content. Such claims often cast a website's use of algorithms either as non-publisher activity to which Section 230(c)(1) does not apply, or as "development" of third-party content that renders the service an "information content provider." Federal courts of appeals that have considered this issue thus far have mostly rejected these theories.⁷⁴

The Second Circuit's decision in Force v. Facebook⁷⁵ and the Ninth Circuit's decision in Gonzalez v. Google⁷⁶ each offer detailed analyses of these theories, as discussed in more detail below. Both Force and Gonzalez involved claims seeking to hold social media platforms liable for terrorist attacks under the Anti-Terrorism Act (ATA), a statute that permits legal recovery against someone who commits or supports the commission of international terrorism.⁷⁷ Each case presented a similar theory: in short, that social media platforms had made friend or content suggestions to users, and these suggestions helped advance the cause of terrorist groups using the platforms.⁷⁸ Though the courts in both Gonzalez and Force held that Section 230(c)(1) protects a social media platform's use of algorithms to make recommendations or suggestions,⁷⁹ partial concurrences and dissents in both cases challenge the reasoning that led to this conclusion.⁸⁰ The Supreme Court vacated the Ninth Circuit's decision in Gonzalez, but the Court did so without disagreeing with or otherwise addressing the Ninth Circuit's Section 230 analysis.⁸¹

Algorithmic Recommendations as Non-Publisher Activity

A frequent point of contention in lawsuits brought against platforms is whether claims based on a platform's algorithmic recommendation of third-party content treat the platform as a publisher or speaker.⁸² As discussed above, website features that determine how content is displayed are more likely to be considered publisher activity protected by Section 230(c)(1). For example, the Ninth Circuit held that a message board would be treated as a "publisher or speaker" by claims challenging the message board's use of algorithms to recommend and notify users of potential topics of interest, which allegedly connected a user with a drug dealer.⁸³

The Second Circuit reached the same conclusion in Force.⁸⁴ The plaintiffs argued their claims did not treat Facebook as a publisher or speaker because Facebook's algorithms matching content with users went beyond publisher activity.⁸⁵ However, two of the judges on the Second Circuit's three-judge panel in Force determined that Facebook's use of algorithms to suggest friends and content to Facebook users was publisher activity immunized under Section 230(c)(1).⁸⁶ The majority analogized Facebook's use of algorithmic suggestion to more traditional publisher activities of "arranging and distributing third-party information," such as placing content on a homepage.⁸⁷ The court concluded any act of "arranging and distributing third-party information," including by way of algorithms, "inherently forms 'connections' and 'matches' among speakers, content, and viewers of content, whether in interactive internet forums or in more traditional media."⁸⁸

The Ninth Circuit reached a similar conclusion in Gonzalez. The Gonzalez plaintiffs had argued that Google was liable for allowing the terrorist group known as ISIS to use and access the video sharing platform YouTube.⁸⁹ The court held that these claims "[sought] to impose liability for allowing ISIS to place content on the YouTube platform" and therefore treated Google as a publisher.⁹⁰

Algorithms as Content Development

Plaintiffs may seek to sidestep Section 230(c)(1) by arguing that algorithmically amplifying content "develops" the content and renders the platform responsible for it.⁹¹ Arguments that a platform is responsible for developing third-party content frequently rely on the material contribution and neutral tools tests articulated in Roommates.

The Roommates court itself cautioned that the use of "an ordinary search engine" should not constitute "development" under Section 230.⁹² Other federal courts hearing claims brought against search engines have agreed. In O'Kroley v. Fastcase, Inc., the Sixth Circuit held that Google's display of allegedly defamatory content in its search results did not "develop" the content.⁹³ The court added that Google's alterations to the content did not "materially contribute" to its unlawfulness.⁹⁴ A federal district court similarly held that a search engine's alleged "manipulation" of search results to promote defamatory content did not develop that content.⁹⁵

Several courts hearing claims premised on the use of algorithms have adopted the "neutral tools" analysis from Roommates. In Marshall's Locksmith Service v. Google, a case involving search engines that automatically converted addresses provided by third parties into "pinpoints" appearing on the search engines' mapping websites, the D.C. Circuit emphasized that the search engines' tools did "not distinguish" between different types of user content.⁹⁶ Instead, the algorithms translated all types of information in the same manner.⁹⁷

The Second Circuit reached a similar conclusion in Force.⁹⁸ The plaintiffs in Force argued Section 230(c)(1) did not apply because Facebook's algorithms helped create or develop terrorist content by directing that content to the site's most interested users.⁹⁹ Looking to both the material contribution and neutral tools tests, the Force majority determined that Facebook's involvement in user content was "neutral."¹⁰⁰ The court observed that Facebook's algorithms matched content to users "based on objective factors applicable to any content" and did not "augment[] terrorist-supporting content primarily on the basis of its subject matter."¹⁰¹ These neutral algorithms were insufficient to render Facebook a developer of the user content.¹⁰² The Ninth Circuit applied this analysis in Gonzalez, reasoning that YouTube's recommendation system, while "more sophisticated than a traditional search engine," was still "neutral" in that it treated terrorist-created content the same as other third-party content.¹⁰³

Judicial Challenges to Section 230's Scope

No court has thus far ruled that a platform may be held liable for using algorithms to recommend content. However, several judges have expressed concern over applying Section 230(c)(1) to recommendation systems. In both Force and Gonzalez, one member of the three-judge panel partially dissented and argued that Section 230 should not bar lawsuits under the ATA against platforms for using algorithms to amplify or recommend terrorist content. In Gonzalez, one of the members of the panel "reluctantly" joined the majority and wrote separately to explain her misgivings about the scope of Section 230(c)(1)'s immunity.

Chief Judge Katzmann challenged the Force majority's reasoning in a partial dissent. According to the chief judge, the claims did not treat Facebook as a publisher, because they were based "not on the content of the information shown but rather on the connections Facebook's algorithms make between individuals."¹⁰⁴ Facebook was not merely publishing content, the dissent argued, but "proactively creating networks of people."¹⁰⁵ Chief Judge Katzmann also argued that Facebook's friend and content suggestion algorithms communicate a message: that Facebook believes the specific individual viewing the suggestions will like the suggested content or be interested in connecting with the suggested person.¹⁰⁶ By analogy, the chief judge opined that Section 230 would not protect a third party that analyzed Facebook user data using an algorithm and then sent users messages recommending particular content.¹⁰⁷

Chief Judge Katzmann cabined the reach of his dissent by observing that the claims in Force are "atypical" in that defendants are liable under the ATA for providing services to terrorist organizations.¹⁰⁸ He suggested his approach to Section 230 would not render Facebook liable for "common torts" like defamation in which Facebook's use of an algorithm to boost or recommend content would be immaterial to the claim.¹⁰⁹ In a defamation claim, "the mere act of publishing . . . creates liability," whereas under the ATA, it was the operation of the algorithms that allegedly provided illegal material support to the terrorist organizations.¹¹⁰

Members of the Ninth Circuit's panel in Gonzalez similarly challenged the reasoning behind extending Section 230's protections to algorithmic amplification. In a concurring opinion, Judge Berzon wrote that if Ninth Circuit precedent did not require otherwise, she would hold that promoting or recommending content is not publisher activity.¹¹¹ Citing favorably to Chief Judge Katzmann's dissent in Force, Judge Berzon concluded that recommendations by algorithm "are well outside the scope of traditional publication," which has never included selecting material to display to each individual reader.¹¹² Instead, platforms using algorithms to recommend content communicate their own messages to users about what content they might like.¹¹³ Despite reaching this conclusion, Judge Berzon determined that Ninth Circuit precedent "squarely and irrefutably" held that recommending content is publisher activity.¹¹⁴ She therefore "reluctantly" joined the majority opinion, but urged the full Ninth Circuit to reconsider its precedent.¹¹⁵

Judge Gould dissented in part, stating that he would hold that a website's use of otherwise "neutral tools" is unprotected by Section 230 if 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."¹¹⁶ Like Judge Berzon, he expressed support for Chief Judge Katzmann's Force dissent.¹¹⁷ Judge Gould also urged the full Ninth Circuit or the Supreme Court to address Section 230's applicability to algorithmic recommendations.¹¹⁸

Considerations for Congress

While few courts have explored Section 230's application to algorithmic recommendations in depth, the decisions in Force and Gonzalez may indicate judicial reluctance to limit the broad reach of Section 230 embraced by Zeran and subsequent decisions. Congress may consider whether the broad immunity currently recognized by courts should apply to algorithmically sorted content or, alternatively, whether certain behavior or content should warrant different treatment under Section 230. Any changes to Section 230's protection may also raise concerns under the First Amendment's Free Speech Clause.

Supreme Court Decisions and Section 230

In the more than 25 years since Section 230 was enacted, the Supreme Court has never decided any cases interpreting the law. The Court agreed to hear a Section 230 case for the first time in Gonzalez. The Supreme Court had the opportunity in Gonzalez either to narrow Section 230's scope as it applies to such recommendations or to ratify the appellate court consensus that algorithmic recommendations are protected by Section 230. Instead, the Court vacated the Ninth Circuit's decision without addressing Section 230.¹¹⁹

Twitter v. Taamneh, a companion case to Gonzalez, focuses on liability under the ATA notwithstanding Section 230.¹²⁰ The claims in Taamneh rely on a similar theory as those in Gonzalez: in short, that the defendants aided or abetted acts of international terrorism by allowing ISIS to recruit individuals and spread their message using social media.¹²¹ The Court held in Taamneh that the claims against social media companies did not give rise to liability under the ATA.¹²² Consequently, the Court held in Gonzalez that its decision in Taamneh foreclosed liability for many of the claims.¹²³ The Court "decline[d] to address the application of § 230" and instead vacated the Ninth Circuit's opinion and remanded the case for reconsideration in light of Taamneh.¹²⁴ As a consequence of the Supreme Court's decision, the Ninth Circuit's decision in Gonzalez is no longer binding precedent in the Ninth Circuit. Because the Supreme Court did not disagree with (or even address) the Ninth Circuit's Section 230 analysis, the Ninth Circuit may choose to reaffirm its analysis on remand.

The outcome in Taamneh offers a reminder that a platform's exposure to legal liability for its recommendations does not depend only on whether the platform receives protection from Section 230. Removing Section 230's protections for certain activity, such as promoting or amplifying content, may not subject a platform to liability if the platform's activity is not legally actionable. In other words, an individual alleging harm due to a recommendation would have to state a legally recognized basis for a lawsuit in order to hold the platform liable.

Legislation

Some Members of the 117^th Congress introduced several bills that would have addressed Section 230's relationship with algorithmically sorted or recommended content. These bills generally would have restricted the availability of Section 230's protections for platforms that "recommend" or "promote" certain content. To date, one of these bills, the DISCOURSE Act, has been reintroduced in the 118^th Congress.¹²⁵

Although these bills take varying approaches, any revisions to Section 230 may raise several interpretative questions. Select issues are considered below.

Relevance of Existing Legal Doctrine

As discussed above, a robust and largely consistent body of caselaw interprets the scope of Section 230's protections. Some approaches from federal appellate decisions, such as those from Zeran and Roommates, have enjoyed widespread application even in state and federal courts where the decisions have no binding precedential effect.¹²⁶ One consideration to revising Section 230 is how any changes to the law will interact with these existing legal frameworks. Proposals that explicitly alter Section 230's application might render certain analytical frameworks wholly or partially inoperative. For example, the DISCOURSE Act would have added provisions to Section 230 explaining that platforms engaged in certain activity should be considered information content providers.¹²⁷ Courts may have concluded that this should supplant the "material contribution" analysis most courts now use to determine whether a platform developed challenged content—or they might have tried to integrate the existing analysis into the new statutory framework. Either way, courts would then have to decide what cases decided under the prior framework could still be relied on under the new framework, and flesh out the meaning of the new statutory provisions. Existing frameworks may remain relevant for proposals that condition Section 230's protections but do not otherwise alter Section 230(c)(1) or its definitions.

Congress could also choose to affirm or disavow expressly any existing judicial interpretations of Section 230. If Congress wishes to enshrine any existing analytical frameworks, it may do so explicitly in the text of Section 230.

Defining the Covered Activity

Many of the proposals from the 117^th Congress focus on "recommendation" or "amplification" of particular content as a basis for limiting Section 230's protection, with some variation in the exact terminology used. Some commentators have suggested that determining whether a platform has "amplified" content may be difficult because of the complexity in assessing the role recommendation systems play in directing content to users.¹²⁸ Defenders of Section 230 have argued that any act of arranging or publishing content may necessarily "amplify" certain content, because some pieces of content will be displayed more prominently than others as a matter of design.¹²⁹ The courts in Force and Gonzalez relied on similar reasoning in holding that a platform's use of algorithms is publisher activity under Section 230(c)(1).¹³⁰ A potential consideration is how or whether a Section 230 carveout might apply to internet search functions, which by design "amplify" content that the platform determines is most responsive to a user's search query, or other functions that promote certain content at a user's request.¹³¹ Some past proposals, such as the Protecting Americans from Dangerous Algorithms Act, explicitly except search functions from their coverage.¹³²

Some of the past proposals further limited their covered activity in other ways, such as requiring a platform to have a particular mental state (for example, the Justice Against Malicious Algorithms Act's limitation to reckless or knowing recommendations)¹³³ or describing with more particularity what types of recommendations are covered (such as the Don't Push My Buttons Act's focus on "automated functions" that rely on user information collected by the platform).¹³⁴ Defendants claiming Section 230(c)(1)'s protection currently need only to satisfy the three criteria discussed above.¹³⁵ Exceptions to Section 230(c)(1) that would create additional criteria—such as by requiring that a defendant demonstrate that it has not acted with a particular mental state, or that it does not use "automated functions" to recommend content—may remove some of the procedural benefits of Section 230 by making it harder for platforms to claim its protections without additional judicial factfinding.¹³⁶ Additionally, proposals that focus on automated functions may create a situation where a recommendation made with human input would be entitled to Section 230's protection when the same recommendation made automatically would not be.

Free Speech Considerations

The Free Speech Clause of the First Amendment to the U.S. Constitution limits the government's ability to regulate speech.¹³⁷ Reforms to Section 230 may raise several free speech concerns generally.¹³⁸ Proposals that make Section 230's protections unavailable for certain algorithmic operations raise at least three questions. One question is whether, if Section 230 is unavailable, hosting or promoting others' speech on the internet is itself protected under the First Amendment. If it is, the First Amendment might restrict liability. The other questions relate more directly to the constitutionality of reform proposals: the second question is whether modifying an existing liability regime raises the same First Amendment concerns as enacting a law that directly prohibits or restricts speech. A final question is, if such a proposal does raise First Amendment concerns, whether withholding Section 230's protections for certain algorithmic operations impacts speech based on its content.

Overview of Free Speech Principles

The Free Speech Clause of the First Amendment limits the government's ability to regulate speech.¹³⁹ Courts have long recognized that this protection may extend beyond written or verbal communication.¹⁴⁰ However, the Free Speech Clause does not provide the same degree of protection for conduct as it does for what the Supreme Court calls "pure speech."¹⁴¹ For conduct to receive First Amendment protection, the conduct must be "expressive."¹⁴² In other words, the person engaging in the conduct must have "[a]n intent to convey a particularized message."¹⁴³

The Free Speech Clause generally prohibits the government from regulating speech based on its content,¹⁴⁴ with limited exceptions for certain narrowly defined categories of "unprotected" speech such as defamation.¹⁴⁵ Laws that restrict or burden protected speech based on its topic or subject matter (content-based laws) are "presumptively unconstitutional" and subject to strict judicial scrutiny.¹⁴⁶ Under strict scrutiny, the government must show that the law is the least restrictive means of serving a compelling governmental interest.¹⁴⁷ It is "rare" for a law to survive this test.¹⁴⁸ Courts treat laws that restrict or burden speech based on the expression of particular views (viewpoint-based laws) as a particularly egregious subset of content-based laws.¹⁴⁹ Unlike other content-based laws, viewpoint-based laws are usually categorically unconstitutional and are not justifiable even under strict scrutiny.¹⁵⁰ One instance in which governments may be able to differentiate among viewpoints is when the government itself is acting as a speaker¹⁵¹ or, similarly, when a government chooses to fund certain activities but not others.¹⁵²

Laws that regulate only the time, place, or manner of speech without regard to its content (content-neutral laws), are subject to a lower bar known as intermediate scrutiny.¹⁵³ The test for time, place, or manner restrictions requires the government to show that the law is "narrowly tailored" to serve a "significant governmental interest" and "leave[s] open ample alternative channels" to communicate that speech.¹⁵⁴

A law is facially content-based if it "applies to particular speech because of the topic discussed or the idea or message expressed."¹⁵⁵ A law that is not facially content-based may still be treated as such if there was a content-discriminatory purpose behind the law.¹⁵⁶ A law that merely requires an enforcer to look to the content of speech is not necessarily content-based. In City of Austin v. Reagan National Advertising of Austin, the Supreme Court held that a city law treating "on-premises" and "off-premises" signs differently was not content-based, even though determining whether a particular sign was on- or off-premises required reference to the message on the sign.¹⁵⁷ As the Court explained, a sign's content mattered under the regulation "only to the extent that it informs the sign's relative location" and the law was therefore akin to a content-neutral time, place, or manner restriction.¹⁵⁸

Just as a facially neutral law may be content-based, such a law may be viewpoint-based if the law favors or disfavors a particular point of view in practice. A law that is aimed at particular speakers, the purpose of which is to suppress a viewpoint associated with those speakers, may be viewpoint-based even if the law makes no reference to particular viewpoints.¹⁵⁹ Similarly, a law that excepts some speakers from its application may suggest that the law is targeting particular viewpoints.¹⁶⁰

Whether Hosting or Promoting Third-Party Speech is Protected Speech

As discussed above, if Congress were to amend Section 230 by creating exceptions for certain algorithmic operations, ICS providers would not necessarily be liable for those algorithmic operations.¹⁶¹ Critically, even if a plaintiff could state a claim under some other source of law, the First Amendment might limit liability for protected speech activity. Thus, the First Amendment might provide some immunity even if Section 230 no longer applies.

A number of lower courts have held that websites' decisions about hosting or presenting third-party content are protected by the First Amendment.¹⁶² For example, one trial court concluded that a lawsuit challenging search engine results was barred by the First Amendment.¹⁶³ The plaintiffs argued that Baidu, a Chinese search engine, had violated federal and state civil rights laws by blocking "from its search results . . . information concerning 'the Democracy movement in China' and related topics."¹⁶⁴ The trial court said these allegations would "hold Baidu liable for, and thus punish Baidu for, a conscious decision to design its search-engine algorithms to favor certain expression on core political subjects."¹⁶⁵ In the court's view, allowing "such a suit to proceed would plainly 'violate[] the fundamental rule of protection under the First Amendment, that a speaker has the autonomy to choose the content of his own message.'"¹⁶⁶

These lower court rulings have built on Supreme Court cases recognizing that "when a private entity provides a forum for speech, . . . . [t]he private entity may . . . exercise editorial discretion over the speech and speakers in the forum."¹⁶⁷ This constitutionally protected "editorial discretion" entails the right to choose what material to host and how to present it.¹⁶⁸ While the Supreme Court has not specifically weighed in on how protections for editorial discretion apply to online platforms for third-party speech, the Court has the opportunity to consider the question in two cases being argued in the Court's October 2023 term.¹⁶⁹ While both cases dispute platforms' ability to restrict third-party speech, any Supreme Court discussion shedding light on the general right of editorial control could nonetheless be relevant to determining protections for hosting or promoting third-party speech.

The cases involve conflicting rulings from the Fifth and Eleventh Circuits on state laws that limit platforms' ability to moderate user content.¹⁷⁰ The Eleventh Circuit, in line with the trial court rulings mentioned above, concluded that when social media platforms "'disclos[e],' 'publish[],' or 'disseminat[e]' information, they engage in 'speech within the meaning of the First Amendment.'"¹⁷¹ The court accordingly ruled unconstitutional portions of a Florida law that unduly "restrict[ed] platforms' ability to speak through content moderation."¹⁷² In contrast, the Fifth Circuit upheld a Texas law that, in the court's description, "generally prohibits large social media platforms from censoring speech based on the viewpoint of its speaker."¹⁷³ That court concluded that social media platforms "exercise virtually no editorial control or judgment," and further ruled that "the Supreme Court's cases do not carve out 'editorial discretion' as a special category of First-Amendment-protected expression."¹⁷⁴

As mentioned, the Supreme Court has granted certiorari in both cases.¹⁷⁵ At least until the Supreme Court weighs in, lower courts outside the Fifth Circuit may be more likely to conclude that at least some sorting and promotion of speech qualifies as constitutionally protected editorial activity.¹⁷⁶

Even if a platform's editorial control over third-party content qualifies as speech under the First Amendment, another unresolved question is whether certain algorithmic processes might fall outside this First Amendment protection. The district court in Zhang, discussed above, held that a search engine's use of algorithms that suppressed certain content was protected First Amendment activity, but the claims in Zhang arose from the search engine's "conscious decision" to disfavor a particular political message.¹⁷⁷ Another district court confronting these issues suggested that search engine rankings would be protected First Amendment activity "no matter the motive."¹⁷⁸

As discussed above, algorithms can generate a range of outputs, some more communicative than others.¹⁷⁹ Where on the spectrum of expressiveness a particular algorithm falls may inform a court's decision of how the First Amendment may apply.¹⁸⁰ Platforms may use algorithms to arrange content without consciously deciding to communicate a particular message. For example, a platform may choose to arrange content in a way that maximizes user engagement with the platform.¹⁸¹ The limited caselaw on how the First Amendment applies to algorithms suggests that using algorithms to arrange or rank speech for any reason may be sufficient to warrant First Amendment protection.¹⁸²

Whether Withholding Section 230 Protection Restricts Speech

One question in a First Amendment analysis is whether the government has infringed on protected speech at all. Section 230 grants a statutory protection against liability for publishing third-party speech. Accordingly, it does not restrict that speech, and arguably makes publishing that speech less burdensome in terms of legal exposure. In certain contexts the Supreme Court has recognized that the government does not violate the First Amendment simply by giving a benefit to certain speakers and "declining to subsidize [others'] First Amendment activities."¹⁸³ Under this theory, Congress's decisions to extend immunity only to certain speakers might not implicate significant First Amendment concerns.

Nonetheless, amendments to Section 230 that extend its protection in some instances, but not others, likely could still implicate free speech concerns.¹⁸⁴ Certain proposals that grant immunity only to particular speakers could transgress a core First Amendment principle that the government "may not deny a benefit to a person on a basis that infringes . . . his interest in freedom of speech."¹⁸⁵ Section 230 could be seen to implicate the First Amendment to the extent it denies immunity to those "who engage in certain forms of speech."¹⁸⁶ Certain types of Section 230 amendments could raise the concern that Congress has "discriminate[d] invidiously in its subsidies in such a way as to '[aim] at the suppression of dangerous ideas.'"¹⁸⁷ These "ideas" could include a platform's choices about what speech to recommend, as discussed above, as well as a user's exercise of speech if a proposal encourages platforms to remove content they otherwise would not remove.¹⁸⁸

Several Supreme Court cases suggest that government may sometimes impose content- or even viewpoint-based conditions on government benefits.¹⁸⁹ A more recent Supreme Court case suggests that these precedents might not extend to a non-monetary government benefit such as Section 230's liability protections.¹⁹⁰ In Matal v. Tam, the Supreme Court struck down a provision of a federal trademark statute prohibiting the registration of certain "disparag[ing]" marks.¹⁹¹ The Court held that this provision violated the First Amendment because it was impermissibly viewpoint-based.¹⁹² In defending the provision, the government argued that trademark registration is a benefit provided by the government and that the government "is not required to subsidize activities that it does not wish to promote."¹⁹³ A plurality of the Court rejected the analogy to the line of cases allowing viewpoint-based distinctions when providing government benefits,¹⁹⁴ reasoning that those cases "all involved cash subsidies or their equivalent."¹⁹⁵ Trademark registration is different, the plurality reasoned, because it does not involve the payment of money by the government to a private party.¹⁹⁶ Those cases could not justify viewpoint discrimination in a "government registration scheme" involving non-monetary benefits.¹⁹⁷ The plurality suggested that a more appropriate analogy might be to government "program[s]" or "limited public forums" for private speech where some "content- and speaker-based restrictions are allowed."¹⁹⁸ Still, the plurality did not resolve whether these cases provided the correct framework because the trademark provision involved viewpoint discrimination, which is "forbidden" even in such forums.¹⁹⁹

As with government subsidies, Section 230 provides a type of benefit to private parties; here, in the form of a statutory liability shield. However, as with trademark registration, Section 230's liability protection does not involve payments from the government to private parties. Thus, under the reasoning of Tam, cases authorizing conditions on cash subsidies may not authorize conditions on Section 230: Congress may not be able to claim that conditioning Section 230's protection is necessary to avoid "subsidizing" speech that it does not support. Because the law struck down in Tam was viewpoint-based and not merely content-based, it is unclear how a court might analyze content-based modifications to Section 230's protection—if, for example, a court might find Section 230's immunity analogous to a "government program" or "limited public forum" where some content-based restrictions may be permitted.²⁰⁰ Tam still makes clear that selectively withholding Section 230's immunity based on speech's content or viewpoint could raise First Amendment concerns even if the law does not directly prohibit speech.²⁰¹

Content-Based vs. Content-Neutral Speech Regulations

Assuming that Section 230's protections are analogous to neither a monetary subsidy nor a limited public forum, another question is whether an exception to Section 230—such as one that prevents service providers from using Section 230 when they have recommended content using an algorithm—would be considered a content-based or content-neutral regulation of speech. As discussed above, content-based laws are rarely constitutional, whereas content-neutral laws are subject to a lower standard of judicial scrutiny.²⁰² While "intermediate scrutiny" is easier to satisfy than strict scrutiny, courts may still strike down a content-neutral law if the law burdens more speech than is necessary to achieve its legislative purpose.²⁰³

A law that removes Section 230's protection for certain types of content, but not others, could be subject to challenges in court as a content-based restriction on speech.²⁰⁴ Some proposals from the 117^th Congress would have taken this approach, such as a bill that would have made Section 230's protections unavailable for a website that uses algorithms to promote "health misinformation."²⁰⁵ Several proposals from the 117^th Congress would have removed Section 230's protection for all claims when a provider uses an algorithm to recommend or amplify content.²⁰⁶ Some commentators have suggested that this approach would more likely be assessed by reviewing courts as "content-neutral" and therefore subject to a more relaxed standard of constitutional scrutiny.²⁰⁷

Although some modifications to Section 230's immunity regime based only on the use of algorithmic recommendation or amplification may be content-neutral with respect to user content, a separate question is whether such a change would be content-neutral with respect to a service providers' speech. As discussed above, courts have generally held that a platform's ranking choices—such as the ordering of search results—are "speech" protected by the First Amendment.²⁰⁸ Determining whether this "speech" has recommended or amplified third-party content requires reference to the content of the platform's ranking choice and therefore may be content-based. As the Supreme Court suggested in City of Austin, a law is content-based when its application depends on the "substantive message" of the material regulated.²⁰⁹ Courts might determine that a law regulating recommendation systems, but not imposing any subject-matter or viewpoint restrictions on the material recommended, does not turn on any substantive message and may be assessed as a content-neutral law under City of Austin.

A similar consideration is whether a law withholding Section 230's protections for algorithmically recommended speech might be viewpoint-based. Even facially neutral laws may be viewpoint-based if they discriminate between viewpoints in operation or are motivated by a discriminatory purpose.²¹⁰ A law that affects "recommendations" by algorithm but not processes that reduce the visibility of content may also be viewpoint-based because it would penalize messages in support of recommended content, but not messages that disfavor that content.²¹¹ A counterargument may be that a law that applies to all algorithmic recommendations, irrespective of the content recommended, does not disfavor any particular viewpoint.²¹²

The Supreme Court has struck down laws that, in the Court's view, target particular speakers for differential treatment, essentially treating these laws as content-based burdens on speech subject to strict constitutional scrutiny.²¹³ Courts may also ask whether a law targeting particular speakers—such as a subset of platforms, like social media platforms—is aimed at suppressing particular viewpoints.²¹⁴ Challengers to Florida's and Texas's social media laws, discussed above,²¹⁵ made these arguments, though both the Eleventh and Fifth Circuits rejected the arguments.²¹⁶ Platforms might also argue that a law that discourages platforms from ordering results or arranging content based on an automated process, but that does not discourage ordering results based on direct human input, discriminates against speakers that use automated processes and is therefore akin to a content-based law.²¹⁷ In Turner Broadcasting System v. FCC, the Supreme Court declined to treat a law targeted at cable television operators as content-based, reasoning that the law at issue distinguished among speakers "based only upon the manner in which speakers transmit their messages . . . and not upon the messages they carry."²¹⁸ Courts might apply this reasoning to any legal provision aimed at algorithmic amplification to determine that such a law is subject only to intermediate scrutiny.

Key Takeaways

Platform liability for recommendation systems is an emerging issue in the American legal system. Federal appellate courts that have addressed the issue have unanimously held that platforms may not be held liable for algorithmically recommending third-party content, based on the robust protections provided by Section 230. The caselaw interpreting Section 230 extends back to the time of the statute's original drafting and interprets statutory language that remains largely unchanged since 1996. If Congress wishes to amend Section 230 to directly address a platform's liability for recommendation algorithms, questions may emerge about how to reconcile Section 230 caselaw with any new language. Amendments to Section 230 might be subject to constitutional challenges, and platforms might also argue that their recommendation choices are protected by the First Amendment.