TikTok After Moody: State Courts and Algorithmic Curation

Alexander Zhao is a student at Harvard Law School. He holds a B.S. in Electrical Engineering & Computer Science and Business Administration from the University of California, Berkeley.

Introduction

The Illinois Attorney General, in October 2024, joined thirteen other states in filing enforcement actions against TikTok for violating state consumer protection statutes and committing common-law torts. [1] The Attorney General’s complaint claimed that the social media company violated the Illinois Consumer Fraud Act by designing an array of platform features that harm young users, compromising their safety and mental wellbeing. [2] In part, the Attorney General claimed that TikTok’s recommender system—the algorithm which suggests videos to users on the platform’s “For You” page—induces excessive and addictive use, and that TikTok misrepresented the addictive nature of its product. [3] At the motion to dismiss stage, Illinois overcame a First Amendment challenge from TikTok, who argued that recommendations made by its algorithms constitute protected speech. [4] Part II-C of the Illinois complaint sought to preempt such concerns by averring that TikTok’s recommender system is not expressive because the algorithm is aimed solely at maximizing users’ engagement with the platform, without any concern for the subject matter of its recommendations. [5]

Taking the complaint’s facts as true, First Amendment protections should not bar Illinois’s suit against TikTok, whatever its merits in other respects. While Moody v. NetChoice treated platforms’ curation of third-party content as protected editorial judgment, the Court expressly reserved the question of whether that protection extends to “algorithms [that] respond solely to how users act online.” [6] TikTok’s algorithm falls squarely into that category. Picking up where the Court left off, state trial courts are correctly holding that this class of recommenders is not presumptively entitled to First Amendment protection by rejecting motions to dismiss. Including Illinois, nearly all fourteen of the states that filed in October 2024 have defeated such motions made by TikTok. [7]

This note describes these state court opinions and presents doctrinal justification for this emerging trend. On Illinois’s facts, TikTok’s purpose is strictly to predict user preferences based on a history of user interactions, without reference to characteristics of the recommended items themselves. No human programmer at TikTok is “disfavoring” any content; the algorithm aims only at user engagement. It does not editorialize, because it does not rank based on any “general feature” of the communication or its creator. [8]

I. Moody and State Courts

Though this note focuses on TikTok, it is far from the only company to deploy recommendation algorithms. Illinois’s case against TikTok awaits trial against a background of mounting losses for large technology companies. In March 2026, a New Mexico jury found that Instagram had recommended potential child abusers’ accounts to minors on the app, holding Meta liable for $375 million. [9] A bench trial on the state’s public nuisance theory is in progress. Though the New Mexico trial was an enforcement action, private plaintiffs have prevailed as well. Just the day after the New Mexico verdict, in a case brought by a plaintiff who alleged the platforms had harmed her mental health as a child, a Los Angeles jury found Meta and YouTube liable for defective product design in a bellwether trial. [10] On June 9, the presiding judge denied Meta and Google’s motion for a new trial, rejecting their Section 230 defense on the grounds that the verdict turned on design features rather than content. [11]

Justice Kagan, writing for the majority in Moody, appeared to construe protections for editorial judgment broadly: “The First Amendment offers protection when an entity engaged in compiling and curating others’ speech into an expressive product . . . is directed to accommodate messages it would prefer to exclude.” [12] She emphasized that Facebook and YouTube’s recommender systems implement their Community Standards and Guidelines, embodying views on what content “the platforms disfavor.” [13] Yet in a footnote, the Court carved out an exception for non-expressive algorithms: “[A]lgorithms [that] respond solely to how users act online—giving them the content they appear to want, without any regard to independent content standards.” [14] The Court expressly did not rule on whether such algorithms were protected; Justice Barrett elaborated on this possible exception in her concurrence. [15]

Since then, lower courts have begun to invoke this exception. In District of Columbia v. Meta, the district court cited the above language from Moody when declining to grant Meta’s motion to dismiss. [16] A New York judge did the same in New York’s case against TikTok, stating: “The Court in Moody . . . emphasized that it was not deciding whether the First Amendment applied to algorithms that display content based on the user’s preferences. . . . [T]he complaint does not seek to hold defendants liable for portraying necessarily their own editorial choices, but rather, as plaintiff alleges, the editorial choices of the user.” [17] Just two weeks later, the Illinois trial court also denied TikTok’s motion to dismiss on similar grounds, writing that Moody “explicitly excluded the type of claim . . . framed within the State’s Complaint.” [18] In North Carolina, the trial judge, citing Moody, wrote plainly: “[T]he complaint supports an inference that a reasonable person would understand TikTok’s video feed to reflect a given user’s content choices as opposed to ByteDance’s own creative expression.” [19]

Beyond the Moody carveout, courts have rejected TikTok’s First Amendment challenges on two additional grounds. First, “allegedly addictive features” like endless scroll ostensibly cause harm “regardless of the nature of the third-party content viewed.” [20] Second, to the extent that an AG alleges fraud, “the First Amendment does not protect knowing misrepresentations in commerce.” [21] Neither rationale directly confronts the issue of whether recommendation is protected speech. Reliance on the carveout reserved by Moody may prove to be more durable on appeal, though its viability remains dependent on factual details of TikTok’s algorithm.

II. TikTok’s Recommender System Is Not an Editor

TikTok is a social media platform owned by Chinese internet company ByteDance. [22] Used by over 150 million Americans, its “For You” page serves short videos to users from other accounts, personalized based on users’ past engagements with TikTok. [23] That recommendation engine is distinct from content moderation. Moderation removes offensive content and embodies a company’s independent content standards; a recommender only adjusts an item’s rank, affecting how likely users are to see it. A social media platform might rely solely on content moderation to handle removing offensive content, while its recommender system prioritizes items only by engagement. As noted by commentators, virtually all content moderation algorithms embody a company’s independent content standards. [24] But the same is not true of recommender systems. [25]

A. Expressiveness as the Test

Recommender systems span a spectrum of expressiveness. In this context, expressiveness refers to the degree to which an algorithm manifests the views or preferences of its creator. Much as the Court stated in Brown v. Entertainment Merchants Ass’n [26], an algorithm is expressive when its outputs transmit ideas from its creator to the end user. Further, expressiveness is a matter of degree, not a binary. The Moody Court held that the amount of editorial judgment is proportional to the strength of First Amendment protections. [27] Consequently, courts must draw a “constitutional line” where an algorithm is sufficiently expressive to merit protection from regulation. [28]

On one extreme of the spectrum, a social media feed that displays items in reverse chronological order is a “non-expressive way” of organizing others’ speech and does not merit First Amendment protection. [29] At the other pole, a newsletter with selections curated by a human editor who makes judgments about each note’s quality is certainly expressive speech, even if she does so with programming tools like spreadsheets or a word processor. Between lies a gradient of expressiveness, where courts must draw a brightline.

Importantly, expressiveness is largely independent of technical sophistication. Neither a reverse chronological timeline nor word processing software used by a human editor is cutting-edge technology, but the two sit on opposite sides of the spectrum. Sophistication, however, poses an evidentiary challenge. Highly complex recommender systems, particularly those relying on deep learning, are notoriously opaque. [30] Like most other recommender systems serving billions of items to millions of users, TikTok’s algorithm leverages deep learning, iteratively re-tunes its parameters frequently, and draws on vast troves of data. [31] What does it mean for such a system to manifest a preference? And how traceable would such a preference be to ByteDance? These questions seem impossible to answer if one examines only the recommendation engine’s outputs.

To streamline the inquiry, courts and plaintiffs have advanced a brightline: Does TikTok’s recommender system make decisions on the basis of features about the videos themselves? If TikTok’s recommender system does not rank videos based on their features, then it follows that there is no human-made content preference being transmitted through the algorithm as a conduit. The algorithm is not making any judgments about the videos. In other words, the umbrella of human decision-making extending to the algorithm is attenuated enough that the recommender is making independent calls. [32]

B. Applying the Brightline

TikTok’s algorithm incorporates deep learning to iteratively improve over time based on training data provided by programmers. [33] Companies designing such “smart” recommenders have two primary levers to shape the outputs of a recommendation engine: the inputs it processes and the objective function it aims to optimize. First, the recommender system intakes data as input, which constitutes the raw material for its recommendation. TikTok’s engineers, by assembling the dataset that the algorithm ingests, exercise substantial control over its decisions after the training phase. That data may be either a feature of the videos, or the interaction history between videos and users. Interaction history covers only the touch points between the video and TikTok users, such as the number of likes, shares, comments, or the viewing history of users who interacted with the video. Video features are intrinsic to an item; length, resolution, and subject matter are analyzable even if the video were stripped of all interaction history. The Illinois Attorney General’s complaint alleges that TikTok’s recommender system harnesses exclusively interaction history, without any weight granted to video features. [34] Such a recommender system thus treats the videos themselves like dark matter: a substance that cannot be perceived but only apprehended through influence exerted on surrounding matter.

The second lever is the recommender system’s objective function, which determines how it uses its input data to improve its outputs. During training and fine-tuning, the recommender maximizes its objective by iteratively optimizing its outputs. This objective represents the task at which the designer wants the model to excel. Though the complaint does not specify precisely TikTok’s objective function, it states that the recommendation engine is “designed and trained to optimize for user engagement.” [35] Conventionally, an engagement-focused recommender seeks to minimize the difference between its rankings and the user’s true, underlying preferences, so as to increase users’ time spent on the app.

On the complaint’s facts, TikTok fails the test for expressiveness. Its objective function demonstrates that TikTok did not intend for the algorithm to maximize anything other than user engagement. Beyond that, it could not have transmitted preferences about what kind of videos should be ranked higher, because the training data did not include any video features. Examining these two aspects of TikTok reveals a lack of expressiveness and sidesteps the evidentiary problem of proving that the algorithm’s preferences are not analogues of its creator’s.

Returning to Moody, TikTok’s recommender system would likely fail to meet the bar for protected speech articulated by the Court. It not only falls into the exception for algorithms that respond “solely to how users act” [36], but lacks expressive characteristics. First, the algorithm does not embody a set of platform guidelines. The Court emphasized that Facebook and YouTube’s recommender systems implement their Community Standards and Community Guidelines, respectively. [37] Though TikTok has a published set of “Community Principles,” including “civility” and “inclusion” [38], the complaint alleges that its recommender focuses exclusively on maximizing user engagement. Second, the Court reasoned that curation by social media platforms is protected when “based on general features of the communication or its creator.” [39] The complaint states that TikTok does not recommend videos based on features of either its videos or the creators who post the videos. It would stretch the Court’s language beyond reason to view a video’s interaction history as a “general feature of the communication.”

C. Addressing Objections

By predicting which videos users want to watch, TikTok’s recommender system may reflect its creators’ preferences for showing users “relevant” content. For instance, Professor Jack Balkin argues that the “algorithms maximize engagement because human beings programmed them to do so; they are furthering human goals in producing an ‘expressive product.’” [40] However, if relevance alone constitutes a message transmissible by algorithm, its proponents have proven too much. Permitting “relevance” to count as a message would extend protection to non-expressive forms of recommendation such as reverse chronology, as long as the creator alleges an intended message of relevance, and many users find the recommendations useful.

Professor Stuart Benjamin argues that the cable operators in Turner I made decisions about which channels to include, at least in part, on the basis of what they thought customers wanted, and those judgments were found protected by the Court. [41] There are two responses. First, a human recommender is not trained on blinkered datasets that exclude all features of the content. Even a decision made predominantly on the basis of a channel’s popularity will still involve inevitable judgments, subconscious or not, about the channel’s subject matter and quality. Second, the preferences of cable operators regarding which channels to include are protected, because they are preferences by humans. Regardless of how those preferences were arrived at, they constitute a judgment made by a human acting in an editorial capacity. Relevance to viewers may have been a rationale for their decisions, but the operators’ motivations were not the message transmitted through the choice of channels. Contrast this situation with TikTok, where the only mandate conveyed by programmers to the recommender system is “show relevant videos.” What formed the rationale for human decisionmakers in Turner I now forms the entirety of the transmitted message.

[1] Unredacted Complaint, People v. TikTok, No. 2024-CH-09302 (Ill. Cir. Ct. Jan. 31, 2025) [hereinafter Complaint].

[2] Id. ¶¶ 68–79.

[3] Id.

[4] People v. TikTok at 1, No. 2024-CH-09302 (Ill. Cir. Ct. June 12, 2025).

[5] Complaint, supra note 1, ¶¶ 96–117.

[6] 144 S. Ct. 2383, 2404 n.5 (2024).

[7] At least ten attorneys general have prevailed against TikTok’s motion to dismiss (or the state equivalent): D.C., South Carolina, Illinois, New York, Mississippi, North Carolina, New Jersey, Kentucky, California, and Massachusetts.

[8] Moody, 144 S. Ct. at 2403.

[9] State v. Meta Platforms, Inc., No. D-101-CV-2023-02838 (N.M. Dist. Ct. Mar. 24, 2026).

[10] Social Media Cases, JCCP No. 5255 (Cal. Super. Ct. Mar. 25, 2026).

[11] Id. (Cal. Super. Ct. June 9, 2026); see also Jody Godoy, Google and Meta Denied New Trial in Youth Social Media Addiction Case, Reuters (June 10, 2026, at 10:26 ET), https://www.reuters.com/world/google-meta-denied-new-trial-youth-social-media-addiction-case-sources-say-2026-06-10/.

[12] Moody, 144 S. Ct. at 2401.

[13] Id. at 2403.

[14] Id. at 2404 n.5.

[15] See also id. at 2410 (Barrett, J., concurring) (“But what if a platform’s algorithm just presents automatically to each user whatever the algorithm thinks the user will like—e.g., content similar to posts with which the user previously engaged? The First Amendment implications . . . might be different for that kind of algorithm.”) (citation omitted).

[16] No. 2023-CAB-6550, 2024 WL 5700129, at *13 (D.C. Super. Ct. Sep. 9, 2024); see also NetChoice v. Bonta, 761 F. Supp. 3d 1202, 1221 (N.D. Cal. 2024) (“When it comes to feeds that recommend posts based solely on prior user activity, there is no apparent message being conveyed.”).

[17] Transcript of May 28, 2025, at 69–70, People v. TikTok (N.Y. Sup. Ct. June 12, 2025) (No. 452749/24).

[18] People v. TikTok at 24, No. 2024-CH-09302 (Ill. Cir. Ct. June 12, 2025).

[19] State v. TikTok Inc., 2025 NCBC 47 (N.C. Super. Ct. Aug. 19, 2025) (citing, inter alia, Moody, 144 S. Ct. at 2404 n.5; Moody, 144 S. Ct. at 2410 (Barrett, J., concurring)).

[20] P.F. v. Meta Platforms, Inc., No. 23SMCV03371, 2025 WL 3207662, at *2 (Cal. Super. Ct. Nov. 5, 2025); see also Vermont v. Meta Platforms, Inc., 2024 WL 3741424, at *6 (Vt. Super. Ct. July 29, 2024) (“Unlike Moody, where the issue was government restrictions on content, as discussed above it is not the substance of the speech that is at issue here.”); Commonwealth v. TikTok, Inc., No. 24-CI-00824 (Ky. Cir. Ct. Feb 20, 2026) (“[T]he Commonwealth’s claims . . . target[] the commercial design and marketing of the product, not the communicative content of individual user posts.”); People v. TikTok, Inc., No. 24CV449203 (Cal. Super. Ct. Mar. 6, 2026) (“[T]he State’s claim pertains to Defendants’ design features, not the specific content provided by Defendants.”); Commonwealth v. TikTok Inc., No. 24-2638-BLS1 (Mass. Super. Ct. May 4, 2026) (citing Commonwealth v. Meta Platforms, Inc., 2024 WL 4648435, at *8–9 (“[T]he Commonwealth’s claims are based on . . . conduct and product design, not on expressive content.”)).

[21] State v. TikTok Inc., No. 2024-CP-40-06018 (S.C. Ct. Comm. Pleas May 6, 2025).

[22] Since January 19, 2025, TikTok has been subject to a nationwide ban, requiring ByteDance to either divest TikTok to an American buyer or cease operations. See Sapna Maheshwari, What We Know About the TikTok Ban, N.Y. Times, Apr. 4, 2025. As of May 2025, the Trump administration has paused the enforcement of the ban for a second time. Even if ByteDance were to divest TikTok or lease its algorithm to an American owner, the same legal challenges with its recommender system would remain pertinent. See Peter Kafka, There’s No TikTok Deal (Yet), Bus. Insider (Apr. 3, 2025), https://www.businessinsider.com/tiktok-deadline-will-it-be-banned-donald-trump-sale-2025-4.

[23] TikTok For You, https://support.tiktok.com/en/... (last visited Apr. 10, 2025).

[24] Evelyn Douek & Genevieve Lakier, Comment, Lochner.com?, 138 Harv. L. Rev. 100, 136 (2025) (arguing that the class of algorithms that fall within Moody’s exception for non-expressive algorithms is “probably close to nil”) (citing Tarleton Gillespie, Custodians of the Internet: Platforms, Content Moderation, and the Hidden Decisions that Shape Social Media 1 (2018) (“[A]ll platforms moderate.”)).

[25] District of Columbia v. Meta, No. 2023-CAB-6550, 2024 WL 5700129, at *13 (D.C. Super. Ct. Sep. 9, 2024) (“So although regulations of community norms and standards sometimes implicate expressive choices, the design features at issue here do not.”).

[26] 564 U.S. 786, 790 (2011).

[27] Moody v. NetChoice, 144 S. Ct. 2383, 2398 (2024) (explaining that “[c]urating a feed and transmitting direct messages . . . involve different levels of editorial choice, so that the one creates an expressive product and the other does not”).

[28] Id.

[29] Id. at 2432 (Alito, J., concurring) (finding that “compilations that organize the speech of others in a non-expressive way (e.g., chronologically) fall ‘beyond the realm of expressi[on]’”) (quoting Hurley v. Irish-Am. Gay, Lesbian, and Bisexual Group of Boston, 515 U.S. 557, 569 (1995)).

[30] Yingqiang Ge, et al., A Survey on Trustworthy Recommender Systems, 3 ACM Trans. Recomm. Systems 13:1, 13:15 (2024).

[31] Although TikTok has not publicly described its recommendation algorithm in detail, it almost certainly utilizes deep learning. See, e.g., Zhuoran Liu et al., Monolith: Real Time Recommendation System with Collisionless Embedding Table (Sep. 27, 2022), https://arxiv.org/pdf/2209.076... (Bytedance researchers delineating a framework for training recommendation models with deep learning); Sara Fischer, Inside TikTok’s Killer Algorithm, Axios (Sep. 10, 2020) (TikTok disclosing its use of machine learning for automated recommendation).

[32] Stuart Minor Benjamin, Debate: Algorithms and Speech, 161 U. Pa. L. Rev. 1445, 1478–79 (2013) (predicting that “a different line is tenable and might do significant work in the future...excluding outputs that do not reflect human decisionmaking”).

[33] See supra note 31.

[34] Complaint, supra note 1, at 21–24.

[35] Id. ¶ 100.

[36] Moody v. NetChoice, 144 S. Ct. 2383, 2404 n.5 (2024).

[37] Id. at 2403.

[38] TikTok Community Principles, https://www.tiktok.com/communi... (last visited Apr. 10, 2025).

[39] Moody, 144 S. Ct. at 2403 (emphasis added).

[40] Jack Balkin, Moody v. NetChoice: The Supreme Court Meets the Free Speech Triangle, 2024 Sup. Ct. Rev. 127, 165–66 (2025); see also Mailyn Fidler, The New Editors: Refining First Amendment Protections for Internet Platforms, 2 J. Emerging Tech. 241, 287 (2021) (“[A] decision to design an algorithm to select for relevancy is a human-made choice about what content to prioritize—relevant content.”) (citing Benjamin, supra note 32).

[41] Benjamin, supra note 32 (citing Reply Brief for Appellants Turner Broad. Sys., Inc. at 19–20, Turner I, 512 U.S. 622 (1994) (No. 93-44), 1993 WL 664649).