Federal Court Rules That Blocking Speech Is Protected By The First Amendment, But Recommending It Isn’t

Content Moderation

from the regulating-the-mirror dept

This ruling came out a week and a half ago at this point and I’ve been so annoyed with it that I kept putting off writing about it, but it’s so bad that it requires a discussion. A federal district court in San Jose denied motions from a bunch of the big tech companies seeking to get a preliminary injunction blocking a California law, SB 976, or “Protecting Our Kids from Social Media Addiction Act.” Back when an earlier version of this bill was introduced, I dug into the justification given for it by California state Senator Nancy Skinner, and was shocked to find that it was just blatantly junk science.

Either way, a later version was passed and signed into law by Governor Gavin Newsom. The law defines “addictive feeds” as basically any algorithmic recommendation system on the internet as long as it has some level of personalization. It then requires that websites with such feeds block them from being used by children unless there is “verifiable” parental consent (good luck verifying that). Even more terrifying is that it includes a mandatory limit on children’s access to so-called “addictive feeds” to just one hour per day by default. The law also gives tremendous power to the Attorney General, who gets to determine by what means a site must determine if someone is a child, which effectively means that the AG can mandate intrusive and privacy-destroying age verification.

This is all silly for many reasons, including that the science doesn’t support the claims of addictive feeds, and research actually suggests that constantly calling them addictive does real damage by convincing people they’re helpless to change their behavior.

The fact that California justifies this law by pointing out the fact that a lot of kids use online services that have recommendation algorithms is meaningless. Similar claims would have applied to TV and video games in past decades. And attempts to regulate children’s access to those have repeatedly been struck down as unconstitutional. And while California pointed to the misleading Surgeon General’s report on kids and social media, they seemed to have skipped over the part where he found that social media was actually quite helpful for many.

The trade association NetChoice sued to block the law. Because of some of the awkward language in the Supreme Court’s Moody v. NetChoice ruling regarding the differences between “facial” and “as-applied” challenges to the law, the court said NetChoice didn’t have standing to bring an “as-applied” challenge. Instead, Meta, YouTube, and TikTok all brought such challenges as themselves, rather than via NetChoice.

Judge Edward Davila has now rejected the preliminary injunction they sought in that case, relying on different awkward language in Moody, in which the majority (in a footnote) says it’s not even considering the First Amendment status of algorithmically recommended content. The majority put a clear stake in the ground: content moderation is editorial discretion, and thus protected by the First Amendment, but left the question of recommendation feeds for a later date. Justice Barrett in a concurrence wondered — based on nothing in particular, and certainly not on a thorough briefing on the question — whether algorithms might magically remove First Amendment protections.

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?

She doesn’t actually explore this or answer it in any real way… just uses it as an excuse to say that the ruling in Moody should be narrowly focused on the type of challenge, rather than the merits.

And yet a few courts have seized on this to say that algorithmic recommendations get no First Amendment protections. And that includes this court, where Judge Davila announced that recommendation feeds are somehow not expressive, thus not editorial discretion, and thus deserve no First Amendment protections:

The Court finds Plaintiffs have failed to show at this stage of litigation that their personalized feeds are “expressive.” Rather, Plaintiffs rely on predictive algorithms that incorporate users’ past watch history and other data and then suggest content that the algorithms anticipate will be engaging, or “interesting” to users. Plaintiffs’ decision to rely on these algorithms’ number-crunching capabilities is not an expressive judgment. Although Plaintiffs’ algorithms differ slightly, each Plaintiff represents that their personalized feeds are the result of multi-step processes intended to convey their desired “expressive message.” In other words, each Plaintiffs’ intended message is some approximation of “we think you will find this content interesting.” For example, TikTok’s feeds convey its message that “this content is consistent with our values and we think you will find it informative and/or entertaining.” TikTok Reply 6–7, ECF No. 68. YouTube’s message is similarly that “the user will likely find the displayed content interesting and enriching.” YouTube Mot. 14. Meta likewise intends to communicate a message to its users that, “their particular interests and preferences are reflected on Facebook, Instagram, and Threads services.” Meta Mot. 13. But, as discussed above, Plaintiffs are not making any decisions about what content will be “interesting,” because they are merely relying on predictive modelling to assess what users’ characteristics and history on the platform suggest will keep these same users engaged. This decision is not an “expressive” message; it is merely a mirror that reflects back to users their own perceived interests.

I’ve seen some in the media pick up on this “merely a mirror that reflects back” language as if it matters. But… huh? Even if it is designed to “reflect back to users their own perceived interests” that’s still… very much an expressive decision. It is the site’s judgment about what content will reflect back to users what they want.

That judgment is an opinion; opinions are expressive by nature, and expression is what the First Amendment protects.

Also… it’s not as if calling it a mirror somehow makes it okay to regulate the output of algorithmic recommendation systems. Mirrors are not, famously, a heavily regulated industry.

But more to the point, plenty of expressive work is built around reflecting back to consumers what the publisher or the store thinks those consumers want. Being a “mirror” to consumers’ interests is… what editorial discretion is quite frequently about.

And, look, I know that many people will have a knee-jerk reaction supporting this ruling simply because they hate these companies and hate algorithms, but you won’t like where this ends up. If algorithmic recommendations aren’t protected speech, then the government gets to dictate what companies can and cannot recommend.

Do you really want the Trump administration dictating what content can and cannot be recommended on the internet?

Even more to the point, algorithmic recommendations are also… search. And, yes, the law technically exempts search, but under what basis? If we truly believe that algorithmic recommendations don’t get First Amendment protections, it won’t protect search, even if this particular law chose to exempt it. Do you really want Trump to be able to tell Google what search results it can and cannot provide? The reasoning here makes no sense at all.

The companies, in their briefs, pointed out (correctly) that recommendations are simply the flip side of content moderation, and the majority opinion in Moody made it clear that such things are clearly protected expression. Here was Justice Kagan in the majority in Moody:

Deciding on the third-party speech that will be included in or excluded from a compilation—and then organizing and presenting the included items—is expressive activity of its own. And that activity results in a distinctive expressive product.

But… isn’t that exactly what the state of California is doing here by telling companies how their recommendation algorithms work? It is interfering with how these companies are “deciding on the third-party speech that will be included or excluded from a compilation—and then organizing and presenting the included items.” Judge Davila says no because recommendations and moderation are somehow magically different, leaning on the weird idea that content moderation involves a “moral valence” but personalization algorithms don’t:

A further examination of the characteristics that set content-moderation decisions apart from mere personalization underscores why the former are expressive and the latter is not. Content moderation decisions, implemented through Plaintiffs’ Community Guidelines, carry with them a moral valence; the same is not true for personalization efforts based on users’ data…. By way of example, Plaintiffs do not decide to show a user a series of cat videos because their top-level decisionmakers think that cats are “good” animals or that something about the cats showcased makes the videos especially fun to watch. In fact, Plaintiffs’ inclusion of this message has nothing to do with whether its boardroom believes cats are good or bad, or holds a view on cats at all for that matter. Rather, Plaintiff has decided, devoid of any expressive judgment, that regardless of the curated content itself, feeds should show content that “will be interesting” to users because they have previously watched and completed videos on a certain topic, typically watch videos of a certain length, or are in a certain location. Indeed, if Plaintiffs made an editorial decision about whether to prioritize cat videos in users’ feeds, the decision would doubtlessly affect some faction of users’ long-term satisfaction with the platform, thereby undercutting Plaintiffs’ objectives.

But… um… what says that the First Amendment only applies to content decisions based on a “moral valence”? I’ve not seen any such First Amendment decisions ever. Indeed, you could argue that tabloid magazines, pulp fiction, pornography, and many other expressive vehicles don’t use a “moral valence” for their editorial decision-making, but are very very much based on what the publishers behind them believe will most excite a potential consumer.

How is that any different?

Davila’s justification for the “moral valence” line is citing Brown v. EMA, which was the case that struck down California’s violent video game law fifteen years ago. He notes:

The First Amendment clearly protects “moral judgments about art and literature.” Brown v. Ent. Merchants Ass’n, 564 U.S. 786, 790 (2011)

But that’s not arguing that non-“moral judgments” get no protection. So it’s bizarre for Davila to read it that way. That case alone should tell him why this law is unconstitutional.

Separately, Judge Davila claims that you can easily separate moderation from recommendation based on gating vs. ranking, leaning on content moderation’s role as a “gate” as the reason it’s considered expressive, while noting that even as a content moderation tool may block “harmful” content, the recommender system might still promote it:

For example, Plaintiffs Community Standards reflect their expressive judgment that certain content, such as content that encourages self-harm is bad. But Plaintiffs recommender systems may nonetheless promote this same content through their algorithmic processes.

But that distinction doesn’t survive contact with how these systems actually work. Downranking is moderation. Every one of these companies runs an entire tier of content policy that removes nothing at all and simply makes disfavored content less likely to be recommended — Meta has called it “reduce” for years, YouTube has a whole published framework for “borderline content” that stays up but doesn’t get suggested. Those are written, human, value-laden judgments about which speech the platform doesn’t want to amplify. They are exactly the “moral valence” Davila says personalization lacks. And they live in the ranker, not the gate.

Davila says this proof that the two systems are different. It actually proves the reverse. When a recommender promotes content that the platform’s own standards say is bad, that’s a failure in the ranker — and the fix, the thing every trust and safety team on earth actually does, is to reach into the ranker and demote it. He’s put his finger on the exact seam where moderation and personalization fuse, and then drawn the constitutional line straight through it.

If the court’s argument here holds, then that would mean it’s constitutionally protected to delete and block information… but that it’s constitutionally unprotected to… promote content. That should be self-evidently backwards. The less speech-restrictive choice would get weaker protection. That just seems fundamentally, obviously, wrong.

A recent paper by Corbin Barthold lays out multiple reasons why AI outputs should absolutely be protected by the First Amendment. He highlights the Moody argument mentioned above that Davila dismisses, but also focuses on the fact that the First Amendment includes a strong “right to read” concept. After listing out eight separate Supreme Court cases that make it clear the First Amendment covers the right to receive even extremely dubious or dangerous information, Barthold explains:

The government may not stop you from receiving information from any crank or fanatic who comes to your door. It may not stop you from receiving information from a foreign adversary. It may not stop you from possessing information that is obscene. It may not stop you from receiving crappy advertisements. Given all that, how could you not have a right to receive information from an LLM?

One might object that you have a right only to receive whatever information is out there—whatever information is left after the government regulates LLMs at the source. But Smith forecloses that move. The bookseller couldn’t be held strictly liable for every book on his shelves, not because he enjoyed some special privilege, but because his burden would become the public’s burden. Force the intermediary to self-censor, and you censor the public. So too here. If the government bans certain outputs, imposes vague duties, or otherwise pressures an AI firm, it restricts the public’s access to constitutionally protected information. The state can’t abridge the right to receive information by choking off the channels through which the information flows.

The companies in this lawsuit made a similar argument which again the court dismisses with barely an explanation:

Because Plaintiffs have not shown that their personalized feeds constitute fully protected speech, the Court likewise finds Meta’s argument that the Act impermissibly burdens its First Amendment right to disseminate fully protected speech is not likely to succeed on the merits.

The court also suggests that there is no burden on the right to receive information, because users can always search for it instead of having it recommended:

In turn, minor users remain free to search for specific content, thereby receiving information and ideas.

But, as noted above, nothing in the principle laid out here protects search. If algorithmic recommendations receive no protections, then the next law can similarly mandate how search results must be ordered. They can use the same arguments: kids use search all the time! The top results can be deemed harmful!

The reasoning here is totally circular. It’s claiming that because it’s decided that the personalization is not protected speech as a compilation… that means there’s no right to receive the underlying content. Which is First Amendment protected. Under that logic, any compilation of protected speech can be deemed unprotected… merely by the act of compiling, thereby eliminating the public’s’ right to receive the underlying protected speech.

How can that possibly be right?

And yes, this is just the denial of a preliminary injunction, and the court admits that upon further briefing it could change its mind regarding the expressive nature of algorithmic recommendations, but this seems like a very ominous start — especially for a law that the Attorney General is allowed to start enforcing immediately. Any kind of suppression of speech is supposed to be considered unacceptable, which is why injunctions are common while courts explore the details. Choosing to let this law be enforced already pretty much guarantees that some expression will be suppressed. Looking later to see whether that suppression was okay under the First Amendment is too late. You’ve already created the irreparable harm.

This ruling will almost certainly be appealed. But if the reasoning survives, it hands the government a power it has never had: the ability to dictate what gets (and doesn’t get!) recommended to you, on the theory that a recommendation loses its expressive value the moment a machine delivers it. In the digital era where everything is intermediated by computers, that would be unprecedented power in the hands of the government to determine what it is most people are likely (and not likely) to see.

Filed Under: 1st amendment, algoirthms, amy coney barrett, editorial discretion, edward davila, free speech, ranking algoirthms, rob bonta
Companies: google, meta, tiktok, youtube