A federal judge in San Jose refused on August 5, 2026 to block three provisions of California Senate Bill 976, holding that the ranking systems behind personalized feeds carry no expressive message of their own and therefore fall outside First Amendment protection.

Judge Edward J. Davila of the United States District Court for the Northern District of California denied preliminary injunction motions brought by Meta Platforms, TikTok Inc., and Google LLC together with YouTube LLC. The 22-page order, docketed as Document 106 in Case No. 5:25-cv-09789-EJD and entered in the parallel actions numbered 5:25-cv-09792-EJD and 5:25-cv-09795-EJD, permits California Attorney General Rob Bonta to enforce sections 27001(a), 27002(b)(2), and 27002(b)(4) of the Protecting Our Kids from Social Media Addiction Act.

The three platforms had argued that compiling and ordering third-party content for each user is itself an act of expression, comparable to a newspaper editor selecting copy. The court did not accept that framing. According to the order, the plaintiffs rely on predictive algorithms that ingest watch history and other signals and then surface material the models anticipate will hold attention, and that reliance "is not an expressive judgment."

What the court allowed California to enforce

Section 27001 prohibits an operator from serving what the statute calls an addictive feed to a minor unless the operator lacks actual knowledge that the user is a minor or has obtained verifiable parental consent. Section 27002(b)(2) requires operators to give parents a mechanism to limit daily access to such a feed, with the default set at one hour per day unless a verified parent changes it. Section 27002(b)(4) requires a mechanism allowing a parent to set a child's default feed to one in which content is not ranked using information provided by or associated with the user, other than age or minor status.

A further deadline sits inside the same section. By January 1, 2027, operators must comply with regulations promulgated by the Attorney General for assessing whether a user is a minor, according to section 27001(a)(1)(B).

The statutory definition is broad. Section 27000.5(a) covers any website, service, or application in which multiple pieces of user-generated media are recommended, selected, or prioritized for display based in whole or in part on information provided by the user or otherwise associated with the user or the device. Seven carve-outs follow, including feeds built on information that is not persistently associated with the user, feeds driven by search terms unrelated to the user, feeds responding to an express and unambiguous request for specific media or a specific creator, sequential playback from the same creator where playback is not automatic, and direct private messages.

The line the court drew between moderation and personalization

The most consequential passage separates two functions that the platforms had described as inseparable. Meta told the court there is no material distinction for First Amendment purposes between content-moderation algorithms and personalization algorithms. The court disagreed, and it did so by walking through each company's own engineering declarations.

Every system described in the record begins by assembling a candidate pool. Content that violates community rules is filtered out before ranking begins. Meta applies a rules-based first stage that removes what its declaration calls "problematic items," a category that includes sexual or objectionable images. YouTube's recommender only considers videos that have not been removed under its Community Guidelines. Material breaching TikTok's Community Guidelines is excluded from the pool eligible for the For You feed, the company's primary personalized surface.

That gating step, the court found, is expressive. Community rules are written by people, carry what the order calls a moral valence, and reflect judgments about which viewpoints deserve promotion. Meta's Community Standards are organized into twenty-three separate policies, each with a stated rationale. YouTube's guidelines treat content promoting voter suppression as creating a risk of egregious harm.

What follows the gate is different. Meta's declaration describes prediction models including p(like), p(comment), and p(video completion), with the letter p denoting the probability of the event occurring. Those outputs are combined and compressed into a single ranking score. The order's own description of p(comment) states that it predicts the likelihood a user will "like" a piece of content, an inconsistency in the text that does not affect the analysis. TikTok's system scores and ranks eligible videos to determine which would be most interesting and informative to a given person. YouTube's Vice President of Engineering described a system built to predict which content will be relevant, enjoyable, and valuable to a particular user, optimized for long-term satisfaction.

Each company characterized the resulting message in similar terms. TikTok said its feeds convey that "this content is consistent with our values and we think you will find it informative and/or entertaining." YouTube framed its message as a prediction that "the user will likely find the displayed content interesting and enriching." Meta said it communicates that users' "particular interests and preferences are reflected on Facebook, Instagram, and Threads services."

The court read all three as variations on the same sentence and found the sentence empty. Ranking systems, the order states, are "designed to evaluate users, not messages," quoting an amicus brief filed by the Electronic Privacy Information Center and a group of law and technology scholars in support of the Attorney General. The resulting feed, in the court's words, "is merely a mirror that reflects back to users their own perceived interests." Citing Justice Alito's concurrence in Moody v. NetChoice, the order concludes that these feeds function as passive receptacles for third-party speech, or as "dumb pipes" that emit what they are fed.

Human involvement did not rescue the argument

The platforms argued that engineers make expressive choices when they build and tune recommender systems, including the choice to optimize for engagement. That argument appears designed to answer Justice Barrett's concern in Moody that automation can attenuate the link between human editorial judgment and machine output.

It did not persuade. Setting an objective to maximize engagement is not, on its own, an expressive message, according to the order. The court quoted the hearing transcript to make the point: platforms decide to convey something to users because they think it is interesting, relevant, or trustworthy, "but [Plaintiffs are] not filling in that blank. And it's the filling in of the blank, what you're actually choosing to convey or not convey that is the exercise of editorial judgment."

Three fallback arguments also failed

Meta, TikTok, and YouTube advanced alternative theories. Each was rejected.

The first invoked users' right to receive information, relying on Stanley v. Georgia and United States v. Playboy Entertainment Group. The court distinguished both. Stanley concerned obscenity, a category outside First Amendment protection, and involved no analysis of expressive character. Playboy involved cable operators who blocked programming for two-thirds of the day. SB 976 removes nothing from the platforms. Minors may still search for and receive any content, and the statute permits platforms to recommend media responsive to a user's affirmative conduct.

Meta separately argued that the Act burdens its right to disseminate fully protected speech, pointing to case law affirming a platform's right to promote views it finds congenial. Because the court had already found that personalization efforts convey no expressive message, it held the resulting compilation is not fully protected speech and the argument collapses. Meta had also read a footnote in Moody as exempting Facebook's feed from the Supreme Court's reservations, on the theory that content moderation runs through it. The order rejects that reading, warning that the decision cannot be read to mean that any time a content-moderation decision forms part of a feed's overall mix, the entire feed becomes expressive.

The third fallback was a vagueness challenge from YouTube directed at the phrase "verified parent." Parent is defined in section 27000.5; verified is not. The court turned to Black's Law Dictionary, the same source YouTube consulted, which defines verify as proving to be true, confirming or establishing truthfulness, authenticating, or substantiating by oath or affidavit. Ordinary meaning, the order holds, gives operators sufficient notice that they must confirm a person's status as a minor's parent before default settings are changed.

How the case reached this point

The dispute has a layered procedural history. NetChoice, a trade association whose members include YouTube and Meta but not TikTok, filed the first challenge in November 2024. In NetChoice v. Bonta the district court declined to enjoin the personalized feed provisions, finding that the association had not shown most or all covered feeds are expressive, while enjoining sections 27002(a), 27002(b)(1), and 27005.

The Ninth Circuit affirmed the associational standing holding, concluding that assessing where each member's feed falls on a spectrum from expressing a platform's unique message to reflecting users' revealed preferences would require the individual members to participate. That ruling closed the trade association route and opened an individual one. Google, YouTube and Meta filed separate suits against Bonta on November 13, 2025, the same day the appellate mandate issued, with TikTok bringing its own action.

Those individual filings supplied what the earlier record lacked: company-specific declarations describing how each feed is assembled. The platforms produced the detail the Ninth Circuit said was missing, and the detail was used against them.

Scale figures on the record

The order records numbers submitted by the parties and by the California Legislature. TikTok estimated more than 170 million monthly users in the United States. YouTube stated that it hosts 20 billion videos, with hundreds of hours uploaded every minute. Meta described billions of people globally using Facebook, Instagram, and Threads.

On the other side of the ledger, the Legislature found that roughly 95 percent of 13- to 17-year-olds report using at least one social media platform, and that more than one-third report using social media constantly. A report from the United States Surgeon General identified concerns including less healthy sleep patterns and negative mental health outcomes.

What the ruling changes for the marketing industry

The commercial consequence is narrower than the constitutional one, but it is real. California minors whose parents exercise the section 27002(b)(4) option will receive feeds ranked without behavioural signals, and the section 27002(b)(2) default caps daily exposure at one hour absent parental modification. Both reduce the volume and the targetability of inventory served to a demographic that already sits behind layered restrictions.

That layering has been accumulating for two years. Instagram launched Teen Accounts in September 2024, then tightened restrictions across Facebook and Messenger in April 2025, a change that confirmed at least 54 million active Teen Accounts globally. Meta aligned teen content filtering with PG-13 movie standards in October 2025, extended the 13+ tier to India in April 2026, and took the content shield global in June 2026. The company also began using AI age estimation to move accounts into teen settings automatically. Each step enlarges the share of inventory carrying teen-specific constraints.

The broader question concerns the legal status of ranking itself. If a court can separate a recommender system into an expressive gate and a non-expressive scoring layer, the scoring layer becomes available for regulation under ordinary standards rather than strict scrutiny. That reasoning is not confined to organic feeds. Ad delivery systems rank inventory using probability models functionally similar to p(like) and p(video completion), and the order's logic offers no obvious reason why an auction ranking score would be treated as expressive when a feed ranking score is not.

Regulators elsewhere have already reached the same architecture from a different direction. The European Commission preliminarily found TikTok in breach of the Digital Services Act over addictive design in February 2026, naming infinite scroll, autoplay, push notifications, and recommender systems. It made comparable preliminary findings against Meta in July 2026, with an expectation that autoplay and infinite scroll be disabled by default. American litigation has moved in parallel, with the Massachusetts Supreme Judicial Court holding in April 2026 that Section 230 does not shield Meta from state design claims, and with jury verdicts in California child harm cases drawing comparisons to tobacco litigation.

California's own courts have shown a pattern of declining emergency relief against platform regulation. A federal judge refused to block Assembly Bill 2013, the AI training data disclosure law, in March 2026 on similar preliminary posture grounds.

One caveat runs through the order. The court states plainly that this is a preliminary ruling and that further factual development could show that provisions not enjoined now are unconstitutional. For the time being, Bonta may enforce the three challenged sections.

Timeline

Summary

Who: Judge Edward J. Davila of the United States District Court for the Northern District of California ruled against Meta Platforms Inc., TikTok Inc., and Google LLC with YouTube LLC, in actions naming California Attorney General Rob Bonta as defendant.

What: The court denied all three motions for a preliminary injunction against sections 27001(a), 27002(b)(2), and 27002(b)(4) of Senate Bill 976, finding the platforms unlikely to succeed on their First Amendment claims because personalized feeds do not convey an expressive message. Fallback arguments on users' right to receive information, the right to disseminate protected speech, and vagueness of the term verified parent were also rejected.

When: The order is dated August 5, 2026. The underlying law was signed in September 2024, the individual suits were filed on November 13, 2025, and an age assurance compliance deadline falls on January 1, 2027.

Where: The San Jose Division of the Northern District of California, in consolidated proceedings numbered 5:25-cv-09792-EJD, 5:25-cv-09789-EJD, and 5:25-cv-09795-EJD.

Why: The court concluded that ranking content by predicted engagement is a statistical operation rather than an editorial one, distinguishing it from content moderation, which reflects written community standards and human value judgments. Because the personalization layer conveys no message of its own, the First Amendment does not bar California from regulating it, and Bonta may enforce the three provisions while the litigation continues.