Is an algorithm speech? (Inside by Advanced Topics in AI Law and Policy, Class #4.3)
The Trial, the Constitution, and the Line Nobody Can Find
When Instagram’s recommendation engine serves you more of what makes you watch, is that an editorial judgment, like a newspaper choosing the front page? Or is it a product function, like a defective steering mechanism that pulls you into oncoming traffic?
The answer determines whether social media platforms can be regulated at all. Nine Supreme Court justices tried to answer it last year and couldn’t agree. And right now, in a courtroom in Los Angeles, twelve jurors are being asked to find the answer the Supreme Court couldn’t.
8:30 AM, Friday. Welcome back to Professor Farahany’s Advanced Topics in AI Law and Policy Class! We’re on Week 4 of class. If you haven’t taken Monday’s class, where we covered the interview results and evidence debate about social media harms, do that first. Then complete Wednesday’s class, which covered Section 230. Today builds directly on this week’s material.
(There’s still time to get fully caught up this semester! Start with Week 1, class 1 (From How AI Works to What AI Does) to gain a foundational on understanding autonomy and the impact of digital technologies on our well being. Then dive into Class 1.2 on What AI Does to Your Thinking, and Class 1.3 on Protecting Autonomy in Law, Take 1. Week 2 shifts to the impact of digital technologies on attention, beginning with our attention audit in class 2 (Can you pay attention?), 2.2 (The attention evidence gap), and 2.3 (The laws that miss the point)). And in Week 3 we looked at Dark Patterns, beginning with Class 3 (20 Clicks to Cancel), 3.2 (Why Dark Patterns Work), and 3.3 (How Law Addresses Dark Patterns)).
Here’s where we are. On Monday, everyone described harm but nobody wanted to sue. The evidence is contested but growing, and Meta’s own research found causal effects the company suppressed. On Wednesday, we looked at Section 230, which has shielded platforms for thirty years, and Lemmon v. Snap, which carved out an exception for design claims. If the harm comes from how the product is built rather than what users post, the manufacturer can be liable.
Today, we face the constitutional wall, and the regulatory experiments, and the trial that’s testing all of it, right now, in real time.
The First Amendment Problem
On Wednesday, I asked you whether the content/design distinction was workable. Most of you probably said something like: “It works for the easy cases but gets blurry for algorithms.”
Platforms know this. And they have a constitutional argument for why that blurriness should protect them.
The argument goes like this: When Facebook decides which posts to show you in your News Feed, and in what order, with what prominence, that’s an editorial judgment. It’s the 21st-century equivalent of a newspaper editor deciding what goes on the front page. And editorial judgment is protected by the First Amendment.
This isn’t a fringe argument. The Supreme Court endorsed a version of it in Moody v. NetChoice (2024), which is the most important platform law case in a generation.
A quick detour through the backstory. After January 6th and the deplatforming of conservative figures, most prominently the suspension of then-President Trump, Republican-controlled legislatures were furious. Texas passed H.B. 20, essentially prohibiting large platforms from censoring users based on viewpoint. Florida passed a similar law.
Both were challenged on First Amendment grounds. The cases went to the Supreme Court.
What the Supreme Court Said
Justice Kagan, writing for the majority, applied a line of First Amendment cases going back to Miami Herald v. Tornillo(1974). In Tornillo, Florida had a right-of-reply statute, where if a newspaper criticized a political candidate, the candidate could demand space to respond. The Court struck it down. Why? Because compelling a newspaper to publish content it would prefer to exclude intrudes on editorial discretion. The choice of what to print is itself protected expression.
Kagan extended this to platforms, based on three principles:
The First Amendment protects entities that compile and curate others’ speech when the government directs them to accommodate messages they’d prefer to exclude.
Second, this protection doesn’t disappear just because a compiler includes most items and excludes only a few.
Third, the government can’t override editorial discretion by asserting an interest in “balancing” the marketplace of ideas.
And, critically, she said this extends to algorithmic curation. Even when an algorithm makes the editorial choices, those choices reflect human-designed content standards. The automation of editorial judgment doesn’t strip it of First Amendment protection.
Do you see the problem?
On Wednesday, we established that Lemmon treats product design as distinct from publishing, so Section 230 doesn’t shield design claims. But Moody suggests that algorithmic curation is editorial discretion, which is speech, and the First Amendment protects it.
If the algorithm that trapped one interviewee in a distressing content loop is speech, can you regulate it? If infinite scroll is an editorial choice about how to present content, can you require an off-switch?
If you answered C, you’ve just arrived at Justice Barrett’s concurrence, which may be the most important paragraph in the entire opinion.
The Crack in the Wall: Barrett’s Distinction
Barrett agreed with the majority but flagged something crucial. She distinguished two types of algorithmic curation.
Type 1: Algorithms that implement human-designed content standards. The platform decides “we want to promote educational content and suppress misinformation,” and the algorithm executes that vision. This is editorial discretion, she says, so the First Amendment applies.
Type 2: Algorithms that “respond solely to how users act online, giving them the content they appear to want, without any regard to independent content standards.” These behavior-responsive algorithms, the ones that simply maximize engagement based on user behavior patterns, might not be expressive. They might not reflect any editorial judgment at all.
Think about the interviewee whose Instagram algorithm locked onto her distress. Was that an editorial judgment reflecting Instagram’s content vision? (”We at Instagram believe you should see more content that makes you anxious!”) Or was it a behavior-responsive system that optimized for engagement, without regard to any content standards?
Justice Alito made a related point, in that he warned against equating “secret” algorithms with newspaper editors. The analogy to Tornillo may not hold when the “editorial judgment” is made by a black-box system that even the platform’s own engineers don’t fully understand.
The Fifth Circuit, on remand, emphasized that answering these questions requires granular, function-by-function facts, not abstract arguments. Is ranking a Facebook news feed the same editorial activity as filtering DMs? Managing a marketplace? Running a payments service? The court said the “online world is variegated and complex,” and figuring out which functions are expressive and which aren’t requires evidence the parties hadn’t yet produced.
Here’s the spectrum as it’s emerging:
Content moderation decisions (removing/promoting specific posts)
Almost certainly speech
Hard to regulate
Algorithmic curation implementing editorial standards
Probably speech
Hard to regulate
Behavior-responsive algorithms (engagement optimization without content standards)
Barrett: maybe not speech
Maybe regulable
Product features independent of content (Speed Filter, “last seen,” like counts)
Lemmon: product design
Regulable
The legal system is trying to draw lines within the platform. The hard work, and the fight that will define technology law for the next decade, is figuring out where speech ends and product begins.
What Regulation Actually Looks Like: Utah’s Experiment
While courts work through these distinctions case by case, states are legislating. Utah’s S.B. 152 is the most comprehensive example, and it’s worth studying because it targets specific mechanisms from the evidence.
Utah built a package of tools. Each one addresses a different causal pathway:
Age verification + parental consent = Gatekeeping (you have to prove you’re old enough, and your parent has to say yes)
Ban on advertising to youth accounts = Cuts the revenue incentive to capture minor attention in the first place
Ban on “targeted or suggested” content for minors = Addresses the strongest evidence channel from the research: exposure to risk-behavior content through algorithmic amplification
Limits on DMs + search invisibility for minors = Stranger contact prevention
Curfew (10:30 PM to 6:30 AM) = targets Sleep displacement, one of the best-documented pathways to harm
Parental access to minor’s account = Monitoring
Private right of action ($2,500 per incident) = Enforcement teeth
Which of these provisions look like conduct regulation (ads, data use, curfew) and which look like restrictions on expressive functions? The suggestion ban of ”must prohibit targeted or suggested groups, services, products, posts, accounts, or users” directly restricts what the algorithm says to users. Under Moody’s editorial discretion framework, that could be a First Amendment problem.
My best guess is that the ad ban and curfew are the safest, because they regulate conduct (selling ads, providing access) rather than expression. The suggestion ban is the most aggressive, and potentially the most impactful, but also the most constitutionally vulnerable.
At the federal level, KOSA, the Kids Online Safety Act, creates a duty of care requiring platforms to prevent and mitigate specific harms to minors, targeting design features like infinite scroll, autoplay, and algorithmic recommendations. It passed the Senate 91-3. But civil liberties groups, including ACLU, EFF, LGBTQ+ advocacy organizations, have raised a serious concern. If platforms must prevent “harm to minors,” who defines harm? Could a state attorney general decide that LGBTQ+ content “harms” minors? Trans health information?
This is a real tension. Some of the most vulnerable young people, like LGBTQ+ youth facing hostile home environments, find community and support through social media. A blunt regulatory instrument could cut them off from that lifeline while trying to protect them from algorithmic harm.
Meanwhile, Australia took the most aggressive approach anywhere, with a flat ban on social media for anyone under 16. No parental consent exception, with fines up to $49.5 million Australian. It took effect less than two months ago and The High Court challenge is pending.
The Trial Happening Right Now
All of this, from the evidence, the Section 230 distinction, the constitutional constraints, the regulatory experiments, converges in a courtroom in Los Angeles.
K.G.M. is a 20-year-old woman suing Meta and Google. She started YouTube at age 6. She had Instagram, Snapchat, and TikTok by 14. Her claim is that the platforms’ design choices, including infinite scroll, autoplay, algorithmic recommendations, like counts, addicted her and caused depression, anxiety, body dysmorphia, self-harm, and suicidal ideation.
TikTok settled on the eve of jury selection. Snap settled the week before. Meta and YouTube remain.
The legal theory tracks Lemmon exactly. This isn’t about what anyone posted. It’s about how the products were built. The plaintiffs allege design defects, the features manufactured to maximize engagement at the cost of user wellbeing.
Judge Carolyn Kuhl ruled that jurors should consider whether design features, like endlessly scrolling feeds, contributed to mental health harms, rather than content alone. That’s Lemmon, applied.
Under the risk-utility test for design defects, the plaintiffs’ argument writes itself: infinite scroll could have an off-switch. Autoplay could default to off. Algorithmic recommendations could be opt-in rather than opt-out. These alternative designs would reduce harm without eliminating the product’s utility.
Mark Zuckerberg is expected to testify. The plaintiffs will present Meta’s internal research, including Project Mercury. The jury will see documents showing that Meta’s own scientists found causal evidence of harm, and the company chose not to publish it or change the product.
This is one of approximately 2,200 consolidated cases. Over 350 families and 250 school districts are suing. Forty-two state attorneys general have filed separate suits against Meta, seeking injunctive relief that could change how platforms actually operate.
The comparison to Big Tobacco isn’t hyperbole. The playbook is the same with internal documents, evidence the company knew about harm, design choices prioritizing profit over safety, a jury asked to hold the manufacturer accountable. The critical difference is that tobacco causes physical disease with clear biomarkers; social media effects are psychological and heterogeneous. And tobacco has no First Amendment protection. Social media does.
The Full Architecture
Let’s stand back and see everything at once.
We started Monday with a gap, with people who describe real harm but can’t articulate a legal claim, and don’t believe one exists.
We now have the architecture:
The evidence is contested but growing. Haidt says it’s abundant and damning. Orben says the effect sizes are small, but the science infrastructure is deliberately broken by the companies, and waiting for certainty is an industry delay strategy. Meta’s own research found causal harm and was suppressed. The Surgeon General says we “cannot conclude it is sufficiently safe.”
Section 230 shields platforms from content claims but, after Lemmon, not from design claims. The content/design distinction is the legal spine of everything happening now.
The First Amendment constrains what regulation is possible. Moody says algorithmic curation is editorial discretion. But Barrett says behavior-responsive algorithms, the ones that just maximize engagement, might not be speech at all. The line between “editorial vision” and “engagement optimization” is where the constitutional battle will be fought.
Regulation is proliferating. Utah targets specific design mechanisms. KOSA creates a duty of care. Australia bans kids outright. Each has vulnerabilities, constitutional, practical, and moral.
Litigation is testing all of it before a jury. Right now. In Los Angeles. With Mark Zuckerberg on the witness list and Meta’s own internal documents in evidence.
Returning to Your Interviewees
Remember Monday? “It’s my responsibility.” “I consented.” “I can’t quantify the harm.” “It’s futile.”
The legal system is slowly building a different frame. When a product is designed to capture your attention, optimized to learn your vulnerabilities, and built to maximize the time you spend in a state the manufacturer knows is associated with depression and anxiety, is the individual really the right locus of responsibility?
One interviewee wished for “some limit on the algorithms.” Another wanted “meaningful control.” Another disabled “last seen” and felt “much more at peace.” Another just wanted a timer.
None of them imagined the legal system could give them what they wanted.
The legal system is trying to prove them wrong. Whether it succeeds turns on a question that sits at the intersection of technology law, constitutional law, products liability, and the philosophy of speech: Where does “speech” end and “product” begin?
Nobody has the answer yet. But in that courtroom in Los Angeles, twelve jurors are about to be asked to find one.
If you answered E, I hear you. This is genuinely hard. The evidence is real but contested. The legal frameworks are powerful but have holes. The constitutional terrain is treacherous. And the policy choices involve real tradeoffs, like protecting children versus preserving access for LGBTQ+ youth, regulating design versus restricting speech, demanding corporate accountability versus respecting individual autonomy.
The question for the semester: If you can’t trust your own preferences, and the science is uncertain, what should law do?
Next week, we will dive into AI chatbots. In advance of that, try spending 15 minutes each day, for 4 consecutive days, talking to a AI Companion Chatbot like Replika, Character.AI or Nomi. But don’t spend more time than that before we meet.
Class dismissed. See you on Monday!
The entire class lecture is above, but if you’d like to support my work or go deeper in your learning, please upgrade to being a “paid subscriber.”
Paid subscribers also get access to class readings packs, discussion questions, bonus content, full archives, virtual chat-based office hours, additional readings, as well as one live Zoom-based class session per semester.
