Recommender Systems Are Publishing

and not mere hosting of user generated content.

What a Recommender System Actually Is

A recommender system decides what a particular person sees, and in what order, without that person having asked for any of it. Everything else on this page follows from that. A search engine answers a question you posed; a recommender answers a question you did not pose, on your behalf, using a model of you that you did not write.

The technique that made this possible is collaborative filtering, prototyped at Xerox PARC in 1992 and automated by the GroupLens system in 1994. Its principle, as Pablo Castells and Dietmar Jannach put it in Recommender Systems: A Primer (February 6, 2023), is "that people can benefit from the experience and discoveries of other people ... in making future choices." Mechanically, a platform builds a matrix of who interacted with what and predicts the missing entries: people like you engaged with this, so you will too. Nothing in that method inspects the content. It never has to.

Modern systems keep the principle and industrialize it in two stages: candidate sourcing narrows an unbounded pool to a shortlist, and a ranking model — a supervised neural network trained on logged behavior — scores every candidate and orders them. The platforms describe this themselves. Meta's own Facebook Feed AI system card (as updated June 24, 2026) says the system "calculates a relevance score for about 500 posts and puts them in order by this score," and Nick Clegg, writing for Meta on June 29, 2023, put it plainly: "Our AI systems predict how valuable a piece of content might be to you, so we can show it to you sooner."

Because Twitter published its ranking code on March 31, 2023 — the twitter/the-algorithm repository was created March 27, 2023, and TechCrunch (Kyle Wiggers, March 31, 2023) reported a pipeline selecting about 1,500 posts from hundreds of millions, five billion times a day — we need not speculate about what these models optimize. Twitter's own documentation for the "heavy ranker" describes a parallel MaskNet — reported by TechCrunch at roughly 48 million parameters — whose outputs are "a set of numbers between 0 and 1, with each output representing the probability that the user will engage with the tweet in a particular way": favorites, retweets, replies, profile clicks, video playthrough, extended conversation, and so on. [TikTok's own explanation of the For You feed (June 18, 2020) is the same idea in plainer words: "A strong indicator of interest, such as whether a user finishes watching a longer video from beginning to end, would receive greater weight than a weak indicator."

Read those objective functions again and notice what is absent. There is no term for whether a claim is true, whether an act is legal, or whether a viewer is a child. The system predicts engagement, and engagement is all it predicts. When a recommender promotes illegal content, it is not malfunctioning. It is reporting, accurately, that the content works to engage the viewers interest.

Ranking Is Not Hosting

Hosting is making a thing available to whoever asks for it. Publishing is choosing what to put in front of someone who did not ask. This site sits on a commercial web host that does not rank, boost, or recommend it; that host is properly immune for what we write, having made no editorial choice about it. A platform that ranks our words against everyone else's and decides which stranger sees them has made an editorial choice. The difference is not one of degree. It is the difference between a printing press and a newspaper.

We did not invent this line. The platforms draw it themselves, in their own defense. When [YouTube announced on January 25, 2019 that it would "begin reducing recommendations of borderline content and content that could misinform users in harmful ways," it added: "This will only affect recommendations of what videos to watch, not whether a video is available on YouTube." That sentence concedes the argument. Availability and recommendation are two separate acts, performed by the same company, severable at will — [YouTube reported on December 3, 2019 that the change produced "a 70% average drop in watch time of this content coming from non-subscribed recommendations in the U.S." A firm that can cut watch time of a disfavored category by seventy percent purely by adjusting what it recommends, while leaving every video hosted and searchable, cannot then insist that recommending and hosting are the same legal act.

Courts have begun to say so. In Anderson v. TikTok, Inc., 116 F.4th 180 (3d Cir. 2024), No. 22-3061, decided August 27, 2024 — brought after ten-year-old Nylah Anderson died attempting a "Blackout Challenge" video that TikTok's algorithm placed on her For You Page — Judge Shwartz wrote for the panel that interactive computer services "are immunized only if they are sued for someone else's expressive activity or content (i.e., third-party speech), but they are not immunized if they are sued for their own expressive activity or content (i.e., first-party speech)." TikTok's algorithm, the court held, "was TikTok's own 'expressive activity,' and thus its first-party speech." The panel drew our distinction almost verbatim in a footnote: had Nylah found the video through search, "TikTok may be viewed more like a repository of third-party content than an affirmative promoter of such content." Rehearing en banc was denied October 23, 2024, the mandate issued October 31, 2024, and Justice Alito granted TikTok until February 20, 2025 to seek certiorari (application No. 24A681).

No petition for certiorari appears on the Supreme Court's docket: application No. 24A681 records the extension and nothing after it. The case itself ended in the district court, whose docket (E.D. Pa. No. 2:22-cv-01849) shows the matter terminated on June 3, 2025. What matters for the law is what survives: the panel opinion is precedential, was never reviewed, and has not been vacated.

That reasoning came from the platforms' own First Amendment victory. In Moody v. NetChoice, LLC, No. 22-277 (decided July 1, 2024)), Justice Kagan wrote that in constructing feeds the platforms "include and exclude, organize and prioritize — and in making millions of those decisions each day, produce their own distinctive compilations of expression." Their choices, she held, "constitute the exercise" of protected "editorial control." That holding is a gift the industry cannot return. A company cannot claim the feed as its own protected speech on Monday and disclaim it as somebody else's on Tuesday. Justice Thomas put it sharply, dissenting from the denial of certiorari in Doe ex rel. Roe v. Snap, Inc., 144 S. Ct. 2493, 2494 (2024) and quoted by the Third Circuit in Anderson: in the platforms' world, "the moment that responsibility could lead to liability, they can disclaim any obligations and enjoy greater protections from suit than nearly any other industry."

The question is open, not settled. The Second Circuit went the other way in Force v. Facebook, Inc., 934 F.3d 53 (2d Cir. 2019), over the dissent in part of Chief Judge Robert A. Katzmann, as did the Ninth Circuit in Dyroff v. Ultimate Software Group, 934 F.3d 1093 (9th Cir. 2019). The Supreme Court took the issue in Gonzalez v. Google LLC, No. 21-1333, and declined to decide it. That live split is exactly why this argument is worth making now. The statute itself is treated at /section-230.

One clarification, since the two halves of our position are sometimes read as contradictory. We hold that these firms should be regulated as common carriers and held liable as publishers, because the two attach to different conduct. Common-carrier duties govern denial of service: a platform of this scale should not be free to cut a lawful speaker off from the public square. Publisher liability governs amplification: what the platform pushes at people who never asked. Carry everyone; answer for what you promote.

Amplification Is the Product

Follow the money and the design stops looking accidental. Meta reported on January 28, 2026 that of $200,966 million in full-year 2025 revenue, $196,175 million — over ninety-seven percent — was advertising. Alphabet reported on February 4, 2026 that Google advertising brought in $82,284 million of $113,828 million in fourth-quarter revenue, YouTube ads alone accounting for $11,383 million of it.

Advertising revenue is impressions times price, and impressions are produced by attention — which these systems measure as engagement, the clicks, replies, reshares and playthroughs the ranking models above are explicitly trained to predict. Commercial interest and training objective therefore point the same way by construction. Amplification is not a side effect of the business; it is the mechanism by which the business earns. No firm moderates its way out of that while the objective function stays where it is, and the objective function is where the revenue is.

Frances Haugen, testifying before the Senate Commerce Subcommittee on Consumer Protection on October 5, 2021, described the loop from the inside: "The dangers of engagement-based ranking are that Facebook knows that content that elicits an extreme reaction from you is more likely to get a click, a comment, a reshare." Her written statement in the same record charges that Facebook "repeatedly encountered conflicts between its own profits and our safety. Facebook consistently resolved those conflicts in favor of its own profits. The result has been a system that amplifies division, extremism, and polarization."

What Engagement Optimization Has Cost

These are not hypotheticals. Each has a docket or a report behind it.

Molly Russell was fourteen when she died on November 21, 2017. Her inquest concluded on September 30, 2022 that she "died from an act of self-harm whilst suffering from depression and the negative effects of on-line content." In his Regulation 28 report of October 13, 2022, sent to Meta, Pinterest, Snap and Twitter among others, H.M. Coroner Andrew Walker found that "the platform operated in such a way using algorithms as to result, in some circumstances, of binge periods of images, video clips and text some of which were selected and provided without Molly requesting them." A British coroner, describing a fatality, reached for the hosting/publishing distinction. Nylah Anderson was ten; the Third Circuit's opinion records that TikTok's algorithm recommended the Blackout Challenge to her and that she died of asphyxiation.

On October 24, 2023 thirty-three state attorneys general sued Meta,in the Northern District of California, with nine more jurisdictions filing separately, alleging that Meta built "a business model focused on maximizing young users' time on its platforms" and employed "harmful and psychologically manipulative platform features while misleading the public about the safety of those features." Haugen had told the Senate two years earlier that engagement-based ranking on Instagram "can lead children from very innocuous topics, like healthy recipes ... to anorexia-promoting content over a very short period of time."

On February 19, 2024 the European Commission opened formal proceedings against TikTok under the Digital Services Act, examining "negative effects stemming from the design of TikTok's system, including algorithmic systems, that may stimulate behavioural addictions and/or create so-called 'rabbit hole effects'." A regulator now treats the ranker itself, not the uploads, as the thing to be assessed. So should the courts.

What We Ask

We do not ask anyone to take content down. We ask that the decision to push content at a specific human being be treated as what it is — an act of publication, performed by the platform, for money — and that the ordinary law of civil liability apply to it. Legacy media answer for human-curated publication decisions; new media must answer for machine-curated ones.

We also call on every platform to open-source its ranking algorithm voluntarily, as Twitter did on March 31, 2023. We are not asking for a disclosure mandate and we would oppose one. We are asking these firms to let their users see what is being done to them, because a person who can see the objective function is a person who can choose against it — and that choice is the only intelligence in this system worth the name.

This article was written with the help of generative AI, please verify critical facts.