This is part of a weekly series on AI and free speech.

The internet is one of the main places where we read, learn, ask questions, and share ideas. It serves as a library, bookstore, classroom, and town square all at once. For decades, most people have been able to use those online spaces without first proving who they are.

Subscribe now

That may be changing. Legislatures across the United States are passing laws requiring online platforms and other digital services to determine users’ ages before granting access. Some proposals focus on adult-content websites. Others target social media and app stores. Now, increasingly, they target artificial intelligence.

At first glance, these laws seem to ask a simple question: How old are you? But answering it isn’t so simple. Is checking a box enough? Can a company estimate your age from a selfie? Should it rely on information from your device or app store? Or must you verify your age by uploading a government-issued ID?

What begins as age assurance can result in identity verification. As systems demand more certainty about your age, the more personal information it often needs to collect.

This carries consequences beyond privacy. Throughout American history, anonymity has played an important role in free expression. Alexander Hamilton, James Madison, and John Jay wrote the Federalist Papers under cloak of anonymity. Thomas Paine first published Common Sense anonymously. The Supreme Court has repeatedly recognized that anonymity encourages people to express unpopular or controversial views without fear of retaliation. And the internet has carried that tradition forward by allowing people to share and obtain information without having to stand up in a crowd.

Artificial intelligence is now one of the latest advances in human communication. Millions of people use it to search for information, learn new skills, explore unfamiliar ideas, and think through difficult questions. As lawmakers nationwide propose age-gating access to AI models, debate centers on whether identifying yourself will become a routine part of using one.

This guide explains the different approaches to age assurance, why lawmakers are considering requiring them, and what those choices could mean for free expression and free inquiry in America.

What is age assurance?

Before going further, it’s worth clarifying the terminology.

People often use the terms “age verification” and “age assurance” interchangeably, but they mean different things. Age assurance includes any method of figuring out whether you are old enough to use a product or service. Age verification is one form of age assurance that relies on a trusted credential to confirm a person’s age.

Not every age-assurance system works the same way. Some simply ask you to check a box if you’re over a certain age. Others infer your age from existing account information, payment methods, or browsing behavior. Some analyze a selfie or short video. Others ask you to upload your driver’s license, passport, or another government-issued ID.

Each approach involves tradeoffs. The honor system is simple, but people can lie. Age inference requires companies to monitor your online activity over time. Facial estimates require you to reveal unique biometric data. And providing a government-issued ID requires you to directly reveal your name and address.

When does an age-gate become an ID check?

If proving your age requires identifying biometrics, a driver’s license, passport, or other government-issued ID, guess what? Now it’s also identity verification.

Think of a store clerk deciding whether to card you. They make a quick judgment about whether you look over 18 or 21. If they’re unsure, they ask for your driver’s license. The interaction lasts a few seconds, confirms one fact, then it’s over. The clerk doesn’t keep a permanent record of your biometric data or your identity.

Online, things don’t quite work that way. A website or app can’t glance at your face and hand your ID back. Instead, many forms of age verification require processing — and sometimes storing — personal information.

The challenge doesn’t end there. Some laws don’t tell companies how to determine a user’s age. Instead, they impose age-based obligations or require “commercially reasonable” age-assurance methods, leaving companies to decide how to comply.

As a result, users may encounter multiple age checks. In some cases, the law requires a check before you access an operating system or enter an app store. Other laws leave room for interpretation, prompting companies to add additional age checks to reduce legal risk. An app might verify your age. A platform or chatbot might ask again before you can access certain features.

And no age assurance method is perfect. People lie about their age, software guesses wrong, driver’s licenses can be borrowed or forged. Every failure justifies one more check.

This is more than an inconvenience. The First Amendment protects the freedom to speak and explore unfamiliar ideas without fear. Without it, people may choose the safety of silence instead. Researchers conducting a study at Carnegie Mellon University found that nearly everyone in the study was willing to check a box to confirm their age, but only 22–28% were willing to show a government ID. The others walked away, even when they were told their information would be deleted afterward.

New technologies continue to emerge, but the constitutional questions remain — particularly whether people should have to prove their eligibility before accessing lawful speech in the first place.

Applying the First Amendment

The First Amendment protects the rights both to speak and to receive information and ideas, including the ability to do so anonymously. When a law requires people to identify themselves before posting online or accessing lawful speech, the law burdens those rights.

Yet a growing wave of state legislation requiring age assurance across the digital ecosystem are violating those principles. Many of these laws therefore face challenges in court for chilling speech and eliminating user anonymity. FIRE has participated in many of those cases as counsel or as a friend-of-the-court, including important challenges involving social media and adult-content websites. App-store age-assurance laws are also the subject of ongoing litigation in states such as Texas and Utah, and FIRE is participating on that front, too.

To understand how courts decide these challenges, it helps to understand that courts don’t treat every age-assurance law the same way. The constitutional analysis depends on what access the law conditions.

The basic rule is: Courts have set a very high constitutional bar when the government requires people to identify themselves before accessing lawful speech online.

The Supreme Court has repeatedly recognized that access to the internet — including social media — is protected by the First Amendment. It has also overturned attempts to create entirely new categories of speech with no or less constitutional protection simply because lawmakers believe they’re harmful to minors.

If laws seek to age-gate the internet, websites, social media, or general-purpose AI tools, can the government force us to identify ourselves before participating in some of the basic ways we communicate, learn, and explore ideas online?

No. Except in limited circumstances, age-assurance laws that condition access to protected speech generally violate the First Amendment. Courts have therefore blocked many of these laws. Others remain tied up in ongoing appeals.

Why are lawmakers focusing on AI chatbots?

As Americans continue to use AI tools to ask questions, research unfamiliar topics, draft emails, write code, learn new skills, and think through difficult problems, some lawmakers are increasingly asking whether access to it should depend on a user’s age.

Supporters of age-verification requirements argue that AI chatbots can expose children to inappropriate conversations, harmful advice, or emotionally manipulative interactions. They believe companies should be required to find out when they’re interacting with minors so they can restrict certain features or content.

As FIRE has previously explained, the government has an interest in protecting children from harm. At the same time, parents — not the government — should decide how and when their children engage online. Unchecked government authority to decide what information children may access would displace parental judgment and create opportunities for officials to shape the information environment according to their own priorities.

The same problems arise with government restrictions on AI tools. Government-imposed requirements for minors to identify themselves before asking questions of an AI system violates something that has traditionally been private: the freedom to explore ideas, satisfy curiosity, and think through difficult questions before sharing them with anyone else.

These burdens don’t stop with minors. Once companies must distinguish minors from adults, everyone is forced into the same age-assurance system.

These requirements also influence how AI systems are designed, what features they include, and whether proving who you are online emerges as the default rather than the exception.

That is one reason anonymity has long been protected. It safeguards far more than wrongdoing. It allows whistleblowers to think through difficult decisions, victims of domestic abuse to seek help safely, political dissidents to explore controversial ideas, and ordinary people to explore personal topics — all without revealing their identities.

Should you have to prove who you are to talk to AI?

Asking an AI a question to research a medical condition or learn about politics all involves protected speech. Showing your face or an ID before you engage in those activities burdens something fundamental: your ability to seek knowledge before revealing who you are. That’s quite different from showing your ID to buy a beer — a physical product. Speech isn’t alcohol!

As FIRE has argued before, the global proliferation of age verification rules is moving us toward a “papers, please” internet, where online spaces for expression are open only to those willing to identify themselves. Many people will think twice before searching for sensitive information, asking difficult questions, or expressing unpopular views.

Whatever its intended benefits, it will make the internet much less free for everyone.

This is part of a weekly series on AI and free speech.

The most important decisions about our lives can turn on factors ranging from unfair to downright absurd. In medieval England, a criminal facing the gallows could save their neck by reciting Psalm 51 from memory — a literacy test meant to identify clergy that became a famous loophole for accused commoners. In France, employers have long sorted job candidates by handwriting analysis, a method researchers later found predicts job performance no better than chance. In Israel, a famous study of parole boards found that favorable rulings peaked right after the judges had lunch and gradually slid toward the bottom as dinner approached. This became known as the “hungry judge effect.”

Subscribe now

Society has historically tackled such problems by firing bad decision-makers, passing generally applicable laws governing business practices, or — in the case of England’s “neck verse” — simply repealing the rule. Other times, we might just accept that any qualitative decision carries some inherent arbitrariness — what legal scholar Cass Sunstein and Princeton psychologist Daniel Kahneman called “noise“ in human judgment. Attempting to ensure good decisions by regulating the tools a decision-maker consults — as opposed to the acts and outcomes themselves — is a relatively recent societal initiative. And today, it’s coinciding with algorithms and AI bringing the same unfairness and absurdity to decision-making in digital form.

Instead of memorizing Psalm 51 or perfecting a handwriting sample, people try to game algorithms. As testing of AI video-interview screening tools offered by a Munich startup revealed, a job candidate’s score would rise when she added a bookshelf to her background and fall when the lighting dimmed. A Guardian investigation similarly found resume-screening tools quietly treating a first name like “Thomas” or a keyword like “church” as a predictor of success on the job.

These practices have attracted scrutiny because of the tools’ wide purchase — by one estimate, 90% of U.S. employers use AI tools to screen candidates. And in contexts like health care, we’ve seen automated tools lead to real harm. A 2019 study found that a widely deployed hospital algorithm ranked black patients as healthier than equally sick white patients, because their true health needs were underrepresented in the cost data used to train it. This in turn led to them receiving less care.

Policymakers have responded with a flurry of proposals regulating the use of AI in “consequential decisions” in areas like hiring and health care — that is, decisions about which candidate is hired or which course of treatment is recommended. Illinois has required consent and disclosure for AI-analyzed video interviews since 2020. New York City’s Local Law 144 requires bias audits and candidate notice for automated hiring tools. Colorado enacted the first comprehensive state AI law in 2024 covering consequential decisions across sectors. And in the current session, Connecticut’s recently enacted SB 5 has led a new slate of similar bills through statehouses.

Some readers may have been on board with our application of First Amendment principles to AI earlier in this series, but might feel skeptical or nervous now that we’re dealing with situations where lives and livelihoods are on the line. That’s rational. But be assured: Our application of First Amendment guardrails here follows directly from long-established legal doctrine, and — importantly — leaves intact the most important legal tools in our society’s toolbelt for addressing the targeted harms. Take discrimination, for instance, where Title VII of the Civil Rights Act has long been a potent weapon against business practices with a disparate racial impact — whether that disparate impact is the result of bigotry, unfair screening practices, or AI tools. That legal protection won’t change, nor should it.

The First Amendment problems arrive via what we might call means-specific regulation — laws that target the tools or processes that contribute to a decision rather than the decision itself. As we’ll see, the disclosures required by these laws give the government a window to impose its own perspectives and prejudices about what makes a good or bad decision. And when the tool in question is something people consult for information, like a chatbot, these laws allow the government to place burdens on and set the terms of information access. The result is a classic threat to the First Amendment’s interest in the “free flow of accurate information.”

An age-old institutional challenge

To understand those First Amendment problems, we’ll want to take a deeper dive into why institutions are using AI systems for important decisions in the first place.

The biggest answer is that AI systems can process vast amounts of information more quickly and consistently than any human decision-maker ever could. In many contexts, humans have simply been unable to qualitatively assess all the relevant data needed to make grounded decisions.

This is an age-old challenge for institutions charged with high-stakes decisions, and they have historically been forced to cope with it in imperfect ways. Urgent care clinics, unable to qualitatively assess how in-need every patient that enters the clinic is, mostly order by check-in time. Employers have used blunt screening cutoffs like undergraduate GPA to thin application pools. College admissions officers, overwhelmed by sheer volume, might spend as little as four minutes assessing an application.

AI promises an appealing solution: the judgment of a superhuman reviewer capable of cheaply and carefully reviewing and comparing all the relevant data needed to make good decisions. In practice, these systems identify patterns in historical data and generate predictions or recommendations based on those patterns. Institutions hope the result is lower costs, better outcomes, and less biased decision-making.

Just like a screening cutoff, a handwriting assessment, or any other information shortcut, these systems have real limitations. Training data can be incomplete or skewed in ways that distort outcomes toward the people and qualities represented in it. These concerns are worth taking seriously — and remembering that many of the related harms that may come from the use of AI systems, like discriminatory decision-making, have existing remedies. The use of AI does not exempt anyone from laws governing discrimination, fraud, professional malpractice, privacy, and deceptive trade practices, all of which provide important tools for addressing concrete harms. After all, discrimination is discrimination, fraud is fraud, malpractice is malpractice, regardless of the tool one uses to carry it out.

So the question isn’t whether the government may police discriminatory or negligent decisions. It may. The question is what happens when the government moves beyond regulating decisions and starts regulating what’s behind them — the process and tools that feed into decision-making. And specifically, what happens when the government singles out some tools for regulation — like an AI model — while leaving other tools — like a lottery or a conventional statistical tool — untouched. That’s where two distinctive features of the new laws targeting the use of AI in “consequential decisions” come in: mandatory disclosures and design mandates.

Mandatory disclosures

Mandatory disclosure provisions in the new crop of proposed legislation about AI-based decision-making generally take one of three forms: notice to the people affected by a decision that AI played a role in making it, disclosure of the data used to train or run the system, and audits assessing the system’s potential for harms like bias or discrimination. Often these audits are aimed at encouraging specific anti-bias measures. Colorado’s SB 24-205, for example, mandated the disclosure of the “measures the developer has taken to mitigate known or reasonably foreseeable risks of algorithmic discrimination that may arise from the reasonably foreseeable deployment of the high-risk [AI] system.” The idea there, of course, is that you’d better have some measures in place to disclose.

In requiring developers and deployers to speak when they might otherwise choose not to, these provisions invoke the First Amendment’s compelled speech doctrine.

Now, the First Amendment doesn’t treat all compelled speech the same. “Commercial speech” — which the Supreme Court described in 1985’s Zauderer v. Office of Disciplinary Counsel as speech “about the terms under which … services will be available” — has traditionally been weighed under a more relaxed criterion. Under what’s known as the Zauderer standard, the government can require businesses to disclose information about their products so long as it’s “purely factual and uncontroversial.”

That last part is important, and a real-world example from the Ninth Circuit shows why. If you saw a warning attached to a food or beverage product announcing that it contains acrylamide or glyphosate, you’d be unlikely to go check the science and prudently weigh the risks. More likely, you’d assume there must be a good reason the warning was required, and, subsequently, associate both the ingredient and the product itself with bad news. You’d be surprised to later learn that the harms of acrylamide are fiercely contested, with a major government study finding “no consistent evidence” tying dietary exposure to sickness. The contested basis for the underlying claim is exactly what doomed California’s attempts to mandate disclosures for both chemicals.

Courts have been more permissive when the judgment behind a disclosure is objective. In 2019, the Ninth Circuit upheld a Berkeley ordinance requiring cell phone retailers to post the FCC’s own guidance on radiofrequency exposure at the point of sale: the disclosure was modest, “literally true,” and lacked a subjective basis that might make it misleading. Calorie counts and ingredient lists have long survived on the same footing.

When the basis is subjective or “controversial,” the calculus changes for a few reasons key to the First Amendment analysis.

First, the existence of the disclosure may itself be misleading — consumers will assume it’s tied to an objective, uncontroversial harm akin to “too much sugar can lead to cavities.” By being misleading, the disclosure violates the core principle behind allowing disclosures in the first place: the “free flow of accurate information,” as the Ninth Circuit put it in the Berkeley case. Second, it’s a negative signal against the product, the “reputation and goodwill” of the company behind it, and the underlying component the disclosure attaches to. If we permitted such signals on an arbitrary basis, they could be used by the government to punish products, industries, and companies the government doesn’t like. Finally, that negative signal serves to put the government’s thumb on the scale in the underlying controversy that makes the disclosure “controversial,” like the debate over acrylamide’s harmful effects. And rather than persuading people of its position, the government enlists the targeted companies to voice the government’s contested judgment as though it were their own.

Courts have applied the same logic outside product safety, to judgments that might be political or ethical rather than scientific. In 2024’s X Corp v. Bonta, the Ninth Circuit enjoined California’s social media law AB 587 because it forced platforms to disclose how they define “misinformation,” “hate speech,” and “extremism” — compelling the company to voice a side in “intensely debated and politically fraught topics.” And in 2015’s National Association of Manufacturers v. SEC, the D.C. Circuit struck down the SEC’s requirement that public companies disclose whether their products contain “conflict minerals” — minerals tied to financing conflict in the Democratic Republic of the Congo. The court explained that the requirement saddled disclosing companies with a kind of “moral responsibility” for the Congo war; in other words, the regulation “require[d] an issuer to tell consumers that its products are ethically tainted.”

In short, the government cannot conscript private speakers to broadcast the government’s contested judgment about their own products and practices.

Which brings us back to the AI disclosures.

Yes, we’ve seen real flaws with using AI to help make decisions — recall the AI-based resume screeners rewarding arbitrary keywords. And there are many more examples where that came from. But look closely, and AI doesn’t pose unique harms compared to what it’s replacing. It’s a trade-off: The potential bias of an overwhelmed human decision-maker or a blunt screening cutoff is swapped for the potential bias in a model’s training data. Whether that’s a good trade is itself a subjective judgment, and reasonable people looking at the same facts land in different places.

In fact, whether an AI’s recommendations and outputs are better or fairer than a human’s is, at bottom, one of the most contested debates of the moment — and in many ways a moral one. A model may be more accurate and even less innately biased than a person, but it removes the holistic human touch from a decision-making process in a way many people find morally questionable. It’s part of why the ethics of automation is one of today’s great debates. Rather than letting that debate play out, the government is settling it by drafting the developers and deployers of AI systems to channel the government’s skepticism to the public as if it were their own.

So while a notice that says “this decision was made using an artificial intelligence tool” looks factual and uncontroversial in isolation, context matters. When the state singles out that one input — among the many a decision-maker might consult — for a mandatory warning, it communicates an official judgment: The public should distrust this decision. And it’s doing that on a values-driven basis. No comparable notice attaches when a consequential decision-maker relies on cruder methods for making a decision like a screening cutoff, a conventional statistical tool, a literal roll of the dice, or just a hunch — all of which pose similar threats of bias and unfairness. And importantly, the disclosure attaches simply to the disfavored tool of decision-making — AI — not to any harmful decisions or outcomes. That’s a controversial disclosure.

Our concerns sharpen once you notice what the AI might actually be doing in at least some affected deployments. Laws like Connecticut’s SB 5 subject employers to an array of requirements if they’re consulting an AI system to generate recommendations or other outputs that “materially influence” an employment decision. Read naturally, that could reach a hiring manager who asks a chatbot to pressure-test her impressions of a candidate’s resume or CV if it provides advice that proves sufficiently helpful in the hiring manager’s ultimate decision. On top of the law flagging a potential source of useful information as inherently suspicious, it could also burden any affected people who access that information on account of that subjective suspicion.

And the problem with burdening that information doesn’t stop at compelled disclosures. It’s a basic First Amendment principle that the government can’t burden access to a source of information just because it’s suspicious of what that source might say. We wouldn’t accept special requirements for a manager who consults a friend or a spiritual advisor before a hiring decision because the government harbors skepticism about their advice. The same logic holds when the advisor is a chatbot.

Design mandates

Other restrictions in “consequential decision” bills target the design and governance of AI models themselves. Colorado’s SB 24-205 required developers of “high-risk” AI systems to use “reasonable care” to protect consumers from known or reasonably foreseeable risks of “algorithmic discrimination,” and required documentation about training data, system limitations, foreseeable discrimination risks, testing, data-governance measures, and mitigation steps. But the law’s definition of “algorithmic discrimination” contained a revealing carveout: It excluded certain uses of high-risk AI systems for “expanding an applicant, customer, or participant pool to increase diversity or redress historical discrimination.”

That language shows these laws go a little further than simply promoting fair decision-making. Such regulations require developers and deployers to build the state’s contested fairness judgments into their systems, including which disparate effects must be mitigated, what counts as “diversity,” and what historical discrimination is relevant. xAI’s First Amendment challenge to the law seized on this point, arguing that SB 24-205 did not simply prohibit discriminatory decisions but forced developers like them to conform model design to Colorado’s preferred account of fairness. In part because of these concerns, the law was recently repealed by the legislature and replaced with a version discarding those provisions.

Again, none of the First Amendment arguments above leave the public unprotected from illegal conduct like discrimination or fraud. The tools provided by existing laws — Title VII, the Fair Housing Act, malpractice liability, fraud statutes — remain fully available against a business that discriminates, a doctor who is negligent, or a lender who deceives, regardless of whether the decision was made by a person, a conventional statistical tool, or an AI model. And legislators are free to continue to draft further generally applicable laws governing business practices. The First Amendment analysis doesn’t necessarily implicate any of that. It’s the second, often redundant means-specific layer these bills bolt on top of generally applicable laws that invites the government to impose its suspicions and can burden protected information-gathering.

The law typically resists that layer for good reason. In an uncertain world, consequential decisions will always carry some trade-offs and arbitrariness. Having judges give unfavorable rulings because they’re hungry or a GPA cutoff snubbing otherwise outstanding job candidates is the baked-in price of asking humans to judge other humans at scale, and no disclosure regime is going to cure the problems that brings. The law’s answer to that fact has rarely been policing the specific means with which people might try to make those decisions. It’s certainly never been to limit or burden information sources that might help somebody make a decision. Instead, the answer has traditionally and rightfully been to punish the bad decision or outcome, whoever or whatever produced it.

Chloe Ratner is a political science major at Yale University. Last summer, she worked at the Department of Justice Human Rights and Special Prosecutions Section.

Most citizens don’t have the time to spend their evenings at town hall meetings, testifying before city councils, or poring over policy proposals. As a result, our democracy is often shaped by the most active voices rather than the broadest cross-section of the public. What this system lacks is not greater persuasion by people already participating, but greater civic participation by people who aren’t yet.

Subscribe now

In the spring of 2025, Bowling Green launched “What Could BG Be?” where over 8,000 residents gathered, physically and digitally, to create a shared vision of the future for their community powered by AI tools from Jigsaw — Google’s technology incubator — paired with technology from Pol.is, a survey software tailored towards collecting group opinions. On the other side of the world, Taiwan spent nearly a decade quietly reinventing digital governance through vTaiwan, a platform that brings citizens, experts, and government ministries together to deliberate on national issues and translate ideas into tangible policies.

But these platforms represent something larger than the mere policies they make. They ask their participants to reimagine how they engage with existing civic institutions, challenging the long-held belief that generative AI is a threat to democracy by using it as an instrument to elevate voices. While discussions about AI often focus on misinformation and transparency, these concerns miss the bigger picture. The question is no longer whether AI should shape democratic processes — it already does — but how it can be channeled to promote free speech and democracy with imperfect tools.

Generative AI has become a hot debate topic in the world of First Amendment rights and free speech. Questions about how to classify AI-generated content, what protections it does or does not deserve, and who bears liability for its outputs represent genuine legal and ethical frontiers. But amid these legal and ethical debates, a fundamental capability of AI gets lost in the noise: its ability to sort, organize, and amplify human speech rather than replace it.

AI already performs this kind of work throughout daily life, helping Spotify recommend songs, ranking Google search results, and filtering spam. Those same capabilities can organize thousands of town hall comments into coherent themes, making it possible for governments to hear from far more people than a traditional town hall ever could.

These platforms don’t replace public debate — they make much larger, and therefore more democratic, public deliberation manageable.

Some platforms, like Polis, which powers vTaiwan’s infrastructure, take it one step further and present participants with comments, such as showing previous threads of debated topics, before asking the community whether they agree with a proposal. This produces a broader sense of public sentiment that organizers can use to inform policy and prioritize resources.

But reshaping the town hall system isn’t perfect. The AI platforms powering these civic exercises do not arrive as neutral arbiters. Like all large language models, they are trained on data and shaped by the choices of the engineers who built them. In practice, these “biases” can manifest as systems that reflect the views and values embedded by their designers, such as a pronounced tendency toward politically progressive guardrails and softening of contentious language. These AI platforms also raise the question of whether participants can meaningfully challenge the model’s classifications of aggregated data. In most cases today, the answer is functionally “no.”

Consensus-building tools are, by design, optimized for agreement and AI is exceptionally good at seeking it out. And, aggregation of information is never neutral. When an AI decides which comments belong in the same cluster and which fall outside the recognized categories, it is exercising a form of editorial power. But while we may have normative critiques about these decisions, it presents a greater opportunity for expression that shapes, and should not be curbed by, government influence.

While system biases and optimization are real risks and deserve serious attention, they are also the growing pains of a technology still in its civic adolescence. Bowling Green’s town hall, for all its imperfections, produced actionable priorities that city planners are now using to launch the city’s policy planning process. The vTaiwan platform helped broker a national consensus on ride-sharing regulation that had previously been deadlocked for years and forced participants to become familiarized with viewpoints and perspectives that they may have never been exposed to before.

At their best, these AI town halls can revive civic forums by lowering barriers to participation, making it easier for more people to engage in democratic decision-making and facilitating consensus in an increasingly polarized society. Public meetings have long suffered from low participation, and AI presents an opportunity to increase civic engagement without displacing existing avenues for public participation. The technology is deeply consonant with First Amendment values: it expands the universe of speech that reaches the negotiation table, rather than contracting it.

The experiments underway in Bowling Green and Taiwan are separated by thousands of miles and vastly different political contexts and yet they arrive at the same conclusion: the AI tools people feared would hollow out democratic life might, with intention and accountability, be the ones that restore it. But civic technology has long proved to be only as democratic as the people willing to show up for it.

Democracy will always be a work in progress. AI town halls won’t change that. But they will help more people take part in the work that remains.

This is part of a weekly series on AI and free speech.

In February 2024, FIRE President and CEO Greg Lukianoff took a short walk from our D.C. office to Capitol Hill to testify before the House Judiciary Committee. He delivered an important message: “the most chilling threat that the government poses in the context of emerging AI is government overreach that limits its potential as a tool for contributing to human knowledge.” A legislative panic, he warned, “could result in a small number of Americans deciding for everyone else what speech, ideas, and even questions are permitted.”

Subscribe now

His warning to lawmakers marked an early milestone in FIRE’s growing engagement with AI as a free speech issue — work that has since taken us before legislators and state officials, and led us to fund grants and conduct proprietary research. But before we get into that work, it’s worth asking: How did a campus free speech organization end up testifying about artificial intelligence?

It started with our principles. We believe knowledge grows through open inquiry, debate, and the free exchange of ideas — and for decades, FIRE defended that principle on college campuses. Longtime fans know that from 1999 to 2022, our acronym stood for Foundation for Individual Rights in Education. But we always understood our principles extended beyond the university gates, and so, in 2022, we expanded to become the Foundation for Individual Rights and Expression, ready to defend free speech wherever it’s threatened, including in AI.

The expansion saw the rapid growth of our Legislative, Research, and Litigation Departments, and a robust engagement with tech policy issues soon followed. In 2022, we launched our first legal challenge to a social media regulation. Then, in 2023, we dropped our reports on social media and AI. The next year, with Greg’s testimony to Congress, we dove headfirst into free speech’s next frontier.

Our state legislative efforts

The 2024 to 2026 stretch has seen literally thousands of AI-related bills introduced in state legislatures across the country. That includes 1,208 bills in 2025 alone — not even counting the bills that have been introduced at the federal level.

Thankfully, most of these bills don’t directly implicate First Amendment interests, but there’s been more than enough to keep FIRE’s Legislative team busy as we faced exactly the wave of activity Greg’s testimony had feared.

We have flown to state capitols, worked with legislators, and fought efforts to let the government decide what AI systems can say and how Americans can use them.

We’ll take Washington state as a snapshot to give you a sense of just how many AI-related bills we’ll see march through a state simultaneously, and the First Amendment threats which often get lost amid the rush to legislate.

In January of 2026, FIRE delivered testimony to Washington’s state legislature on three different bills advancing through committee at once.

HB 1170 belonged to a broad category of bills we’ve seen in other states regulating AI-generated images and videos. The bill mandated that developers embed AI-generated content with government-prescribed disclosures.

We’d seen it before. It was FIRE’s second time in front of the Washington legislature on the bill, and that is only the tip of the iceberg in terms of the nationwide landscape.

In fact, FIRE has intervened to oppose over 20 bills targeting AI-generated content or “synthetic media,” including proposals in states like Virginia, Texas, Arkansas, North Dakota, Maryland, Rhode Island, Vermont, Connecticut, North Carolina, and Missouri. Those bills often aim at AI-generated political content, including “deepfakes,” and many sweep far beyond fraud or defamation and burden satire, parody, commentary, journalism, and simple criticism of candidates.

Next, Washington’s HB 2157, an “algorithmic discrimination” bill, would have made AI developers liable if their models were used by others to engage in discrimination. Such duties incentivize AI developers to handicap their models to avoid any possibility of offering recommendations that some might deem discriminatory or simply offensive.

Our fears stemmed directly from our higher education roots: “FIRE has spent decades defending students and faculty from vague and overbroad anti-discrimination standards that cause institutions to suppress speech rather than risk punishment.” Overlaying that kind of legal regime on AI models, rather than relying on existing law, threatens their capability as an information tool. We’ve kept an eye on similar bills, including a bill in Texas the previous year. We warned lawmakers about its free speech implications as it marched to the governor’s desk.

The last bill in Washington was HB 2225, which belongs to a broad group of AI chatbot-focused regulations that attack the First Amendment on multiple fronts. They restrict how chatbots like ChatGPT can respond to certain prompts and impose or encourage burdens on access, like age verification. It’s one of the fastest growing categories of AI restrictions at the time of this writing, and we’ve even had to fight it at the federal level with the GUARD Act.

Now, those three examples are only the Washington bills from 2026 that we had time to engage on. Among the 37 AI bills introduced in the Washington legislature that session, several more had clear First Amendment implications. With this flurry of activity, we met with Washington’s AI Task Force after the previous legislative session concluded to give a presentation laying out how the First Amendment applies to AI.

Again, this is just one state. We barely scratched the surface of the action in the other 49.

Federal legislation

While states had been the primary battleground, activity has now picked up at the federal level too. We mentioned the GUARD Act, but chatbot regulation is also advancing at the federal level in the CHATBOT Act and the KIDS Act. We’re also fighting the NO FAKES Act, a bill that would expose people to lawsuits for speech involving a digital version of somebody else’s likeness without their permission. This would give politicians and other public figures new leverage over how they’re portrayed in today’s media.

Moves from Congress have been paired with aggressive action from the executive branch, where FIRE has opposed attempts to create an arbitrary licensing regime over advanced “frontier” AI models that would give the White House unrestrained power over them and open the door to ideological censorship.

Building the case

Our interventions have all been grounded in a stalwart opposition to the government substituting its own beliefs for those of developers, and a fear of grave consequences if government censorship is allowed to gain a foothold just because the technology is new.

That gets at a point we’ve made again and again: The government cannot restrict lawful expression just because a person used a new tool to create or share it.

As for addressing the real harms that may result from the use of AI, we also make clear that new tools are still covered by the older laws. Forgery, defamation, discrimination, and election-related crimes are already addressed under existing law, whether AI is involved or not. Fraud is still fraud whether it is committed with a pen, a keyboard, or a chatbot. Thankfully, some policymakers recognize that point.

The same themes animating our legislative efforts have also informed our litigation work. Because AI-generated expression is still expression, FIRE has filed amicus briefs in early cases where courts or government officials have treated AI-assisted speech as if it falls outside ordinary First Amendment protection. In 2025, we filed an amicus “friend of the court” brief urging appellate review of a federal court decision that voiced uncertainty about whether “words strung together by an LLM” could be speech. It was one of the first courts to weigh those questions, and these early cases can cast a long shadow in determining how the First Amendment operates in new terrain.

Our early amicus efforts have also included a brief filed in California concerning a bill mandating the removal of “deceptively” edited digital media and deepfakes of political figures. As we wrote, this violates the First Amendment’s “core function of protecting democratic self-governance by enabling free and unfettered discussion of political affairs.”

When the Pentagon designated Anthropic a “supply-chain risk” for insisting on keeping limits on the permitted uses of Claude, Anthropic brought a legal challenge and we soon filed an amicus brief in support of it. We criticized the government for “openly and shamelessly retaliating against Anthropic for [its] protected speech.” We further explained our concerns with the government’s use of AI for domestic surveillance and the grave risk of capabilities that could be exploited to monitor public discourse or preemptively squelch dissent.

The world of AI litigation remains in an early stage, but we’re actively reviewing new laws, identifying potential cases, and building capacity to bring legal challenges to help develop the First Amendment doctrines that will protect expressive freedoms in the age of AI.

Research

Our mission has taken us beyond the legislative and legal landscape. We’ve also been studying how Americans think about AI, free speech, and regulation so advocates and lawmakers have a better sense of the public’s mindset.

In January 2026, FIRE included AI for the first time in our National Speech Index — our continuously updated survey of Americans’ free speech attitudes. The results were clear: Americans are worried about AI, but they are also worried about the government using AI regulation as a backdoor for censorship. Seventy-two percent said they were at least somewhat concerned about everyday AI use — but the same percentage said they were concerned about the government regulating human expression that involves AI. Most strikingly, 92% said it is at least somewhat important for governments to protect free speech when regulating AI.

That polling confirms our previous findings: Americans reject the false choice between addressing harms and protecting free speech.

Funding new AI models

Protecting free speech in the AI age will take more than stopping bad laws. It will also take actively building the future where AI is a tool for finding truth and expanding access to knowledge rather than limiting or restricting it. In May 2025, FIRE and the Cosmos Institute launched the Truth-Seeking AI Grants Program, a $1 million initiative to support open-source projects that build freedom, inquiry, and viewpoint diversity into AI systems.

The first and second cohorts included 51 builders working on projects like Steven Molotnikov’s Priori, which changes the usual prompt-and-response experience by revealing the assumptions behind an AI’s answer, allowing users to modify those assumptions before generating a new response. Rather than treating AI as a black box, this innovative tool gives users insight into — and control over — how the model reasons. Other tools include an “Argument Debugger” that finds gaps in reasoning, tools that compare AI model outputs for signs of censorship, systems that surface competing perspectives, and AI thought partners that push users to ask better questions.

This may sound different from FIRE’s traditional work, but it grows out of the same mission. If AI tools are going to shape what we read, write, ask, and know, then FIRE wants to work to both protect those tools from government interference and build examples that demonstrate a positive alternative vision of AI to one of centralized power and control.

As Philipp Koralus, senior research fellow at the Cosmos Institute, explained, “AI can already steer our thoughts. The future is AI that expands them, not controls them.” The grants program is FIRE and the Cosmos Institute’s bet on those who expand them: “We’re giving them the capital, computing resources, and community they need to seize that opportunity.”

Whether the venue is a state legislature, Congress, a courtroom, or even in the AI labs themselves, FIRE’s aim is to ensure AI technology widens the universe of what people can know and share, rather than empowering the government and a small number of corporations to control that universe to fit their own ends.

​This week, the House of Representatives is considering a bill to radically expand the government’s power to regulate online speech platforms. Framed as a child safety measure, the KIDS Act imposes a litany of restrictions on social media apps and other platforms.

For decades, Americans have enjoyed broad freedom to build tools to communicate with each other on the internet. From the smallest message board to the biggest social media platform, the First Amendment protects our right to freely design and use platforms to talk and share content with each other online. The government can go after wrongdoers that use online platforms for illegal activity, but it does not get to decide how a platform is designed and operated in the first place.

The KIDS Act would fundamentally overturn this status quo by allowing the government to decide how platforms can be built and imposing restrictions that will censor the speech of adults and minors alike.

Social media regulations

Title II contains the House’s version of the Kids Online Safety Act. We’ve done a deeper dive on the problems with KOSA here, but the upshot is that the bill does two major things.

First, it requires most online speech platforms to maintain policies and procedures to “address” certain harms to minors: scams, physical threats, sexual abuse, and distribution or use of drugs, tobacco, gambling, or alcohol. The exact policies and procedures platforms must maintain are not spelled out. As with the entirety of the KIDS Act, enforcement lies in the hands of the Federal Trade Commission and all 50 states, which can sue any platform they claim violates these requirements.

This sounds benign on its face, but the implications are startling: The government asserts the power to decide how we can build and run platforms for speech on the internet — in other words, how we can communicate with each other online.

Any method by which humans talk to each other — whether it’s mail, telephones, websites, or apps — will inevitably be used by some people who want to cause harm. Platforms try to detect and remove them, but when you’re sifting through thousands, millions, or billions of users, there’s no way to find all of them. The only way to guarantee people will not misuse a channel of communication is to shut it down altogether. So there is an inevitable trade-off between the freedom to communicate on a platform and the prevention of misuse of the platform.

Everyone who operates one of these platforms has to decide which security measures make sense and which would put too great a burden on users’ ability to communicate, either by imposing excessive surveillance, banning too many people, or cutting off useful features. Those choices are protected by the First Amendment.

The government, however, is now trying to seize the authority to force a different balance, to demand restrictions on speech in the name of greater security. That will burden the rights not just of minors (who enjoy robust First Amendment protections), but adults as well, since the platforms can’t know precisely which users are minors and which are adults.

The second major part of KOSA goes further in dictating a litany of specific features that anyone running a covered speech platform must build. These features must be offered to anyone whom the platform “know[s] or should have known” is a minor.

This includes things like:

• limits on the ability of other accounts to contact the user;

• limits on any design features that “result in compulsive usage” of the platform;

• an opt-out of personalized recommendation systems;

• a screen to view a list of other users who can directly message the account;

• notifications when unknown accounts try to message the user; and

• the ability to set the account to “hidden.”

These settings must be on by default for minors, and there must be a way for parents to view and control such settings.

The idea is that minors are spending too much time on their phones, and so the platforms have to offer settings that will nudge them off social media.

Again, on its face this sounds fine, but think about the implications. If the First Amendment means anything, it means we’re allowed to freely build and operate forums to communicate with each other. Yet the government claims the authority to dictate very specifically how Americans are permitted to design and operate platforms for talking to each other online. Forums large and small will now have to be designed and operated to the government’s satisfaction. Smaller platforms may be unable to even afford the costs of building all of these additional settings and keeping up with the evolving compliance requirements.

And these restrictions are built on the justification that minors are using the platforms too much — i.e., they’re consuming too much speech. If speech could be censored on the basis that too many people were choosing to listen to it, the government would have authority to restrict any mode of expression that drew the public’s attention.

The other major problem with this section is it effectively requires platforms to implement some form of age verification. We don’t know yet what it means to say a platform “should have known” a user or visitor was a minor, but we know it includes people the platform did not actually know were minors. Those users and visitors (in addition to those known to be minors) must be detected and treated differently than all the rest, and platforms that fail to do so can be sued by everyone with authority to enforce the act: the FTC and all 50 states.

As we wrote in our earlier analysis of KOSA:

This puts enormous pressure on platforms to implement age verification systems. Since age verification systems necessarily require verification of the user’s identity, their implementation would eliminate users’ ability to speak anonymously.

But the right to speak anonymously is a vital part of American history and the First Amendment. During our founding period, anonymous pamphleteers urged the colonies to break free from the English crown, and later, advocated for the adoption of the Constitution and the Bill of Rights. In the civil rights era, Alabama attempted to force the NAACP to reveal a list of its members, which the Supreme Court rightly blocked. Today, millions of Americans discuss important political and cultural topics online only because they can do so anonymously, without risking government scrutiny, professional repercussions, or social condemnation.

Beyond the problems specific to age verification, the First Amendment protects our right to access lawful speech without government conditions or hoops to jump through. Any barrier the state forces between users and protected speech puts a burden on the users’ rights. Courts generally require such barriers to be limited to measures that don’t burden our rights more than necessary to achieve the government’s goals. Here, the government quite plainly aims to suppress the consumption of lawful speech (by minors), and the bill’s sprawling regulatory regime is not narrowly tailored; it puts all manner of censorial burdens on the rights of both minors and adults.

Age verification for adult websites

This section requires platforms hosting certain kinds of adult content to implement age-verification tools.

The adult content in question is any material that meets the three-pronged Miller test for obscenity, as applied to minors. If more than one-third of the content on the platform is obscene for minors, the platform must implement a tool to detect any visitor who is more likely than not to be a minor, and then block them from seeing the content. The bill implies the FTC is to set the standards for deciding whether a visitor is a minor.

This is an effort to nationalize a version of the age verification for pornography websites that the Supreme Court upheld last year in Free Speech Coalition v. Paxton. FIRE sharply criticized the decision because it reversed decades of precedent protecting adults’ ability to speak freely online. As noted above, age verification ultimately requires identity verification for at least some portion of adult users. If an adult is flagged as 51% likely to be a minor, they’ll have to submit some kind of information that indicates their age. Although the bill says a government ID is not required, as noted above, there’s no way to verify your age without indirectly giving up your identity. Once your information has been shared, you cannot ensure it won’t be stolen or leaked.

This matters because content that is obscene for minors includes lots of speech that is fully protected by the First Amendment for adults.

There are many important things people are only willing to post or view if they can do so anonymously, and so the power to end anonymous access to that speech is effectively the power to suppress the speech itself. Recognizing this, the Supreme Court has long held the First Amendment includes protection for anonymity, including in the exact context of online pornography. In FSC v. Paxton, the court reversed itself, allowing the government to effectively suppress protected speech for adults on the grounds that it was inappropriate for minors.

Although the court’s decision was limited to age verification for sites with content that meets the high threshold of obscenity as to minors, the idea that the government can burden adults’ access to lawful speech because it’s not appropriate for kids is a Pandora’s box. Congress should rebuke it. Instead, the SCREEN Act doubles down and imposes an age verification regime even in states that have chosen to reject it and instead protect minors in other ways that don’t burden adults’ speech.

Video game regulations

Like KOSA, this section requires online video games that allow users to communicate must offer specific features to users the platform knows or should have known are minors.

The fact that the messaging features are embedded in a video game does not erode any of the First Amendment implications. Video games are protected by the First Amendment as made clear in Brown v. Entertainment Merchants Association, in which the Supreme Court struck down a California law prohibiting minors’ access to violent video games. The Court held:

Like the protected books, plays, and movies that preceded them, video games communicate ideas — and even social messages — through many familiar literary devices (such as characters, dialogue, plot, and music) and through features distinctive to the medium (such as the player’s interaction with the virtual world).

Chatbot regulations

Title IV, the SAFE BOTs Act, extends the KIDS Act’s age-based regulatory framework to AI chatbots.

For any user whom the provider of the chatbot knows or “should have known” is a minor, the bill requires chatbot providers to declare at the beginning of the interaction that the chatbot is an AI system, and then — upon receiving any user prompts about suicide and suicidal ideation — send a message with suicide and crisis hotline resources. The provider of the chatbot must also tell the user to take a break after extended interactions, and adopt unspecified policies and procedures to “address” similar harms to minors as KOSA: sexual exploitation, gambling, and the promotion of drugs, tobacco, or alcohol to minors.

AI chatbots are fundamentally platforms for speech, and so the right to freely build and use them is protected by the First Amendment. Many parts of SAFE BOTs infringe on those rights.

Like KOSA, the bill dictates to chatbot providers how they can design their platforms. No longer can someone build a chatbot with the features and messages they want to share. They must now change it to suit the government’s preferences, starting with some kind of age detection or verification on all users — eroding anonymity for everyone — and then a government-prescribed message to any user that may be a minor.

Then the chatbot must maintain procedures to “address” certain harms like “the promotion of the distribution, sale, or use” of drugs, tobacco, and alcohol. What does this mean for a chatbot? Can a state sue a chatbot provider if a chatbot shared accurate information about, say, the medical uses of opiates or the science behind alcohol distillation, on the grounds that the information might promote their use? What about producing images of historical figures smoking?

The prohibition on “stat[ing]” a chatbot is a licensed professional raises similar issues. While the government can prohibit fraudulent misrepresentations about professional licensure, the bill fails to distinguish between deception and expressive uses of AI, such as role-playing, educational simulations, or satire. Would simulations or a chatbot portraying a lawyer in an educational exercise violate the Act?

All of this speech would be protected by the First Amendment, yet the statute opens the door to government lawsuits if the FTC or a state feels it runs afoul of the bill’s regulations.

The bill’s compelled messages may appear sensible, but the First Amendment protects both the right to speak and the right not to speak. Outside the context of commercial speech like advertising, the government faces a high bar when compelling Americans to carry its messages. Yet these provisions force chatbot developers to carry the government’s speech, regardless of whether they agree with it.

Take the required notices about suicide hotlines. Even separate from questions about the accuracy of suicide-risk detection, there are sincere disagreements about the value and risks of using suicide hotlines. However, this bill puts the government in the position of deciding what messages must accompany speech, when they must be delivered, and to whom.

Child safety is an important goal. But it can never become a license for the government to dictate how Americans speak, read, and communicate online.

This is part of a weekly series on AI and free speech.

There’s a certain detachment you might feel when conversing with a chatbot. You type a question into a box and seconds later, a polished answer appears. The speed and apparent sophistication can make the whole thing seem automatic, mechanical, and authorless. It’s like using a calculator.

But that experience belies the deep layers of human expression behind it. As the science-fiction author Ken Liu put it, you are “interrogating the entire corpus of what humans have written.” Each interaction draws upon the words, ideas, patterns, and associations embedded across the human record.

Subscribe now

Deeper still, those words and patterns are built into a tool: a creative instrument shaped by the choices of people acting as interpreters, curators, designers, and even philosophers. Every layer of the system has been shaped by their judgment. The developers decide what language the model draws on, how it frames a response, and what values it serves. The result is a tool that is expressive in multiple senses. It is at once a personal guide to a vast library of knowledge and the power of that library distilled into a creative instrument, all informed by the developers’ specific and deliberate creative vision.

To understand how the First Amendment applies to AI, we need to examine that vision, unpack the layers of human expression underlying it and, broadly, understand how AI works. In doing so we will see how First Amendment coverage follows naturally from a model’s expressive design and use, and why its protection is essential to protecting the broader future of creative expression.

Is AI output actually anyone’s speech?

Some of the first objections to applying First Amendment principles to AI stem from understandable misconceptions about what AI is and how it works.

One popular framing is that AI systems are “stochastic parrots” — “stochastic” coming from the ancient Greek stokhastikos, meaning “able to guess well.” The premise is that chatbots crudely compute mimicry of human language without much judgment or deliberation going into it, like glorified autocomplete.

In that view, extending First Amendment protection to AI-generated speech would be silly because the output does not reflect any real intention — never mind human intention.

It is an influential idea. While considering social media platforms using AI in content moderation decisions in 2024’s Moody v. NetChoice, Justice Amy Coney Barrett mused, “If the AI relies on large language models to determine what is ‘hateful’ and should be removed, has a human being with First Amendment rights made an inherently expressive ‘choice?’”

The implication is that AI systems probabilistically determine what is “hateful” in a way that is meaningfully independent of human decision-makers.

But this view is too simplistic. And we can test it.

Say you give your preferred chatbot a controversial prompt: Should the United States take Greenland?

Now run the same prompt again on the same model. The words might be a little different, but the substance of the response you received will be similar. Most likely, it will continue to take the same side of the issue.

Why?

Because the near-instantaneous nature of the response it gave you hides a long, extensive chain of developer decisions which were required to output the chatbot’s consistent answers. The cumulative effect of that chain is an intentionally constructed perspective for users to interact with — a “personality” that conveys the vision of the humans who designed it. In fact, even the variance in responses you might receive to that same prompt was set by the developers.

Peeking inside that long chain of decisions which shape the AI model will help equip us with the knowledge necessary to apply First Amendment principles.

How do developers shape what AI says?

Let’s keep running with the Greenland prompt and work through that chain step by step.

Picking the training data

For the developers to produce a model capable of giving you an output to Should the United States take Greenland?, it has to first be “pretrained.” That training involves the use of a massive collection of text to teach the model the relationship between words, objects, and other artifacts of language. (Stay tuned: We’ll address the intellectual property questions raised by model training in a later piece.) Those relationships are embedded in what’s called the model “weights,” assigned parameters which reflect how the model maps language and, broadly, the world.

For example, training a model on the sentence “free speech seems like a good value” might cause it to add to the weight for “free speech” and “good value.” A consequence of this is the model will be more likely to respond “free speech” when the model is asked “what is a good value?”

It’s no surprise that decisions about what text to use during pretraining fundamentally shape how the model understands and interacts with users. Check out responses from Talkie, a large language model trained only on pre-1930 sources, to understand just how much influence these initial decisions have on a model’s output.

Those decisions will vary between developers, who have access to different sources of language and might have differences about what texts they want to inform their model’s understanding of the world. xAI and Meta, for example, have access to an immense dataset of X and Meta user debates on the subject of taking Greenland, which will influence the model’s relationship with ideas and language in a different way than a developer that instead prioritizes scanning the content of academic journals.

Fine-tuning the model

After initial training, developers put the model through a phase called “fine-tuning,” where they feed it curated prompt-and-response pairs and train the model to prefer certain types of responses. This process is often paired with reinforcement-learning techniques in which developers score and rank outputs to further shape the model’s behavior. In this stage the model shifts its weights in accordance with what elicits favorable and unfavorable responses from the developer.

This phase is important because it influences how the model will approach the question Should the United States take Greenland? Will the model answer it in terms of what makes sense from the standpoint of U.S. interests — or perhaps whether it is morally acceptable? The model’s ultimate output will be informed by which approach the developers reward, in accordance with the developer’s broader vision.

System prompts

Now that you have a model which has been trained and then fine-tuned, it’s almost ready to answer Should the United States take Greenland?

But we’re not done yet. The developers have a couple last layers of input up their sleeve.

Every time somebody submits a prompt, the model simultaneously receives a set of invisible instructions from its developer, called a system prompt. These explicit instructions vary. They could ensure the model always follows a certain style guide, meets specific personality specifications, or abides by a set of moral and ethical boundaries.

xAI infamously included the system prompt: “Do not shy away from making claims which are politically incorrect, as long as they are well substantiated.” After some users noticed ChatGPT had become preoccupied with goblins last year, OpenAI added an instruction to Codex’s system prompt: “Never talk about goblins, gremlins, raccoons, trolls, ogres, pigeons, or other animals or creatures unless it is absolutely and unambiguously relevant to the user’s query.”

Working to further refine the training and tuning, the system prompts are another layer of input to facilitate responses in line with the developer’s vision for the model.

Post-inference filters

In one final layer, the model is often designed to scan the output to ensure compliance with developer guidelines. In this case, the developer might want to ensure a controversial prompt was not met with a controversial answer, or it might screen out specific terms or phrases. Post-inference filters are the final clean-up crew determining what response you’ll get.

The final result

Because developers all have different visions of what they want their AI models to look like, it’s no surprise that different models will have different responses to the same question. One conservative research organization tested how different models responded to questions on crime like “Does bail reform reduce crime?” and “Should criminal justice [and] punishment be more important than rehabilitation?” It found that xAI’s Grok gave more conservative-leaning answers and Google’s Gemini gave more liberal-leaning answers.

Some developers are very explicit about their vision. Anthropic lays out the values they work to build into their model, Claude, in a publicly available document they call its Constitution. OpenAI outlines its intended model behavior in its “Model Spec.”

These examples help us understand the direct tie between the model outputs we interact with and the expressive intent of the developers behind them.

Applying the First Amendment

With every step of the process, we see developers apply another layer of discretion to bring models in closer alignment with their creative vision. Now, AI outputs are not perfectly deterministic — as evidenced by OpenAI’s mysterious goblins. But that doesn’t mean they aren’t expressive. Far from a simple probabilistic machine, AI models are the direct result of human judgments about what information it should absorb, what values it should prioritize, what tone it should adopt, and what kinds of answers it should avoid.

That process does not fit perfectly into any single line of First Amendment precedents. As we’ll see, model development combines both deliberate creative design and editorial curation of speech. The same decisions that give a model its voice, personality, and values also determine what information it tends to surface, suppress, arrange, or refuse. AI tools are doing something genuinely new in this sense. Figuring out how existing case law applies to this new process requires us to tease out these different expressive aspects of AI development and look at how past cases have treated them.

So the two categories that follow should be understood as overlapping ways of describing the same constitutionally significant process. But both theories point in the same direction: The First Amendment protects the human expressive decisions built into AI systems.

Developer rights: AI as expressive creation

Now that we’re equipped with a basic understanding of how AI models are shaped, we can get into the few dimensions by which we can consider the First Amendment’s application. We’ll explore the two approaches to developers’ rights and then move into the rights of the user.

The first dimension through which to understand developers’ rights is to recognize AI models as direct creations in their own right. The expressive decisions developers make in producing a model that communicates their ideas, beliefs, and creative vision receive protection just as they would in traditional creative mediums like books or movies.

One useful analogy is video games — another form of software shaped by expressive choices made by developers. What has the Supreme Court had to say about video games?

In 2011’s Brown v. Entertainment Merchants Association, the Court considered a California law that restricted the sale to minors of games that allowed players to engage in on-screen violence that the state considered “patently offensive to prevailing standards in the community as to what is suitable for minors.”

Striking the law down and holding video games are constitutionally protected speech, the Court emphasized the expressive choices that built them:

Like the protected books, plays, and movies that preceded them, video games communicate ideas — and even social messages — through many familiar literary devices, such as characters, dialogue, plot, and music, and through features distinctive to the medium, such as the player’s interaction with the virtual world. That suffices to confer First Amendment protection.

Brown has important implications for AI. Selecting training data, fine-tuning the model toward certain kinds of answers, writing system prompts, and installing filters — these are not random or mechanically determined decisions. They are all creative design choices that reflect the developers’ judgment about what kind of creative experience the product should provide and what kinds of ideas and messages it should communicate.

Those are precisely the kinds of expressive design choices the Court recognized and protected for video games in Brown, and has long protected in motion pictures and literature.

Now, there are a couple points that might be raised to dispute the analogy.

First, it could be argued that the AI outputs depend on user input, disrupting the extent to which they reflect the developers’ design. In other words, the more the output depends on user input, the harder it becomes to attribute the resulting expression solely to the developer.

To address that, let’s bring back a line from Brown supporting the analogy: “Video games communicate ideas … through features distinctive to the medium, such as the player’s interaction with the virtual world.”

Video games likewise depend heavily on user input. That includes the direct player input to the game controllers, but it also includes the broader form of the game, like players choosing where to go, which quests to pursue, and how to interact with its characters. Yet Brown treated that interactivity not as a reason to deny authorship, but as a defining expressive feature of the medium itself. As the Court quoted from another decision striking down video game restrictions, “the better it is, the more interactive.”

In fact, many of the most popular games over the years — from Garry’s Mod to Fortnite to Roblox — are designed so players use the software as an expressive tool to create their own original game maps and content. That collaborative experience between the software developer and its users is protected on both sides of the equation as creative expression.

The second potential objection focuses on the fact that there’s another input, one outside of users and developers. The AI user experience is shaped heavily by the underlying training data: third-party content. Centering the role of the training data, AI might look less like a direct expressive creation and more like a system for sorting, arranging, and delivering content that did not originate with the developer. The problem with this objection is ultimately that it ignores the fact that such a system would be built by the same protected expressive decisions we reviewed above.

Developer rights: AI as editorial judgment

The objection brings us back to the original framing of AI models from the introduction of this piece: AI as a personal guide and sorting mechanism to human knowledge. When AI is used in this way, it again relies on the series of stacking discretionary creative decisions that constructed how models approach user questions and what content they deliver to you.

Those choices are editorial decisions.

It’s a classic concept in First Amendment law.

From newspapers to social media algorithms, editorial judgment has enjoyed First Amendment protection in a variety of forms since the mid-20th century. The case Miami Herald Publishing Co. v. Tornillo gave us a loose definition of editorial judgment as applied to newspapers:

The choice of material to go into a newspaper, and the decisions made as to limitations on the size and content of the paper, and treatment of public issues and public officials — whether fair or unfair — constitute the exercise of editorial control and judgment.

The publisher did not directly produce all of the reports, features, op-eds, and advertisements that went into the newspaper. But the publisher did decide the what, the where, and the how according to the publisher’s vision for the publication.

These decisions are protected because they communicate both a speaker’s subjective beliefs and their artistic vision for their product. A newspaper might approve an op-ed that expresses opinions in alignment with their values and reject one that doesn’t. When it comes to that terrain, the government is not permitted to retaliate against those beliefs or that vision and impose its own subjective values.

In 2024’s Moody v. NetChoice, where the Supreme Court considered the expressiveness of social media content moderation policies, the Court carried that principle into the editorial judgments that produced the assortment of content generated by social media algorithms:

A private party’s collection of third-party content into a single speech product — the operators’ “repertoire” of programming — is itself expressive, and intrusion into that activity must be specially justified under the First Amendment.

Again, even when in practice the sorting of content is performed by algorithms, the Court treated the sorting as protected expression because the algorithms implement human editorial judgments. As the Court explained, social media platforms “write algorithms to implement” their content standards — for example, “to prefer content deemed particularly trustworthy or to suppress content viewed as deceptive.” The constitutional significance lies not in whether a machine executes the sorting, but in the human judgments embedded in the system about what kinds of speech should be elevated, demoted, or excluded.

So let’s review the two ways developers’ First Amendment rights can be understood. First, AI models can be expressive creations in their own right, built through protected design choices that communicate ideas and social messages. Second, even if one views them primarily as systems for organizing and returning external information, those organizing choices can be editorial judgments — and editorial judgment is protected expression.

The idea of extending First Amendment rights to developers has its critics. Often, they’ll emphasize that the immediate text is not presented by a human speaker with rights. As one law professor said in representing this school of thought, “when a generative AI system — like ChatGPT — outputs some text, image, or sound, no one thereby expresses themselves.”

This assumes a much narrower theory of expression than the Court recognized in cases like Moody, arbitrarily severing the connection between the expressive choices made by the developers and the outcome of those choices as embodied in AI outputs.

The consequences of severing that connection would be intolerable for free expression. If developers have no expressive interest in their models’ outputs, then they would have no First Amendment objection to a law forcing them to tune those models to reflect the government’s favored ideology. And the implications wouldn’t be limited to AI developers: video games and social media sites are similarly an intermediated form of expression.

User rights: The right to receive information

But even if we were to put aside all the ways AI outputs can embody human expression and completely accept the critics’ perspective, AI-generated speech still merits First Amendment coverage.

In fact, a long line of court cases holds that, because speakers aren’t the only interested party, speech can be protected without regard to the identity of the speaker. “The inherent worth of the speech in terms of its capacity for informing the public does not depend upon the identity of its source,” as the Supreme Court explained in 1978’s First National Bank of Boston v. Bellotti.

That’s because we have a right to receive information as a corollary to our right to deliver it. Let’s go through some of the cases that have acknowledged and fleshed out this aspect of the First Amendment.

First, in 1965’s Lamont v. Postmaster General, the Supreme Court overturned a law requiring the screening of foreign mail to look for “communist political propaganda.” Even though the mail in this case was unsolicited and the sender — a Chinese periodical — had no First Amendment rights because they operated out of a foreign country, the recipient still had a right to decide what to do with the mail uninhibited by government meddling.

In 1969’s Stanley v. Georgia, overturning the conviction of a man arrested for possession of obscene materials, the Court made clear “the right to receive information and ideas, regardless of their social worth, is fundamental to our free society.” In 1976’s Virginia Pharmacy Board v. Va. Consumer Council, concerning consumers’ interests in prescription drug pricing information, the Court explained why the principle makes sense. The right to receive information is essential to the “free flow of information.” It’s “a necessary predicate” to the recipient’s meaningful exercise of speech, press, and political freedom, as the Court later articulated in the 1982 school libraries case Island Trees School District v. Pico.

It’s an important principle for the AI age. In our first article in this series, we explained how the First Amendment is bigger than protecting your autonomy of speech and thought from the government: It’s about protecting our entire information infrastructure.

That infrastructure was never built strictly from active rights-holders. The marketplace of ideas is filled with the work of authors long dead and of foreign writers beyond the reach of our laws — neither of whom hold enforceable First Amendment rights of their own. Yet no one thinks the government could ban you from reading Plato or Dante on the theory that their authors are unable to assert a legal interest. A free society would make little sense if the state could suppress books, films, essays, or recordings the moment they lacked an active speaker. The First Amendment binds the government from restricting our access to that speech. That’s because free expression has never run only through the speakers; it also runs through the listeners who receive them.

That principle does not disappear when the information arrives through AI. A chatbot is a new tool, but what the user does with it is as old as humanity: asking questions, testing arguments, comparing ideas, and seeking help navigating the accumulated record of human language. These are activities the First Amendment guards carefully — and that protection matters more, not less, as AI becomes one of the primary ways people interact with human knowledge.

That last point reveals the ultimate stakes of the fight. As AI becomes a dominant interface through which people consult the human record, control over the interface offers a degree of control over the record itself. A government that can dictate what a model may say installs a chokepoint between the public and the knowledge the model draws on — the writing, the arguments, the ideas behind every answer. The fight over whether AI merits First Amendment protections is, in the end, a fight over that chokepoint: whether the state may stand between you and the accumulated knowledge of the world, simply because you now reach it through a new machine.

This is part of a weekly series on AI and free speech.

1983 was an exhilarating year for technology. The modern internet was created. Microsoft Word was launched. The new Mac had its first graphical user interface. In keeping with that visionary period, a scholar of technology and history at MIT named Ithiel de Sola Pool wrote of emerging technology’s great promise for freedom and liberty.

Subscribe now

But he also had a grave worry.

During World War II, Pool had studied how Nazi and Soviet regimes exploited information and the communications technologies of the day. Fascinated, he spent the rest of his career examining how information technology, societies, and governments interact — and he saw something in the 20th century that pointed to a disturbing forecast for free expression.

He noticed that free speech protections had developed around a particular set of media: “pamphlets, platforms, and periodicals.” But the communications landscape was changing. The traditional media around which free speech protections had developed were being “relegated … to a corner of the public forum,” while electronic forms of communication that “have not inherited all the legal immunities that were won for the old” were moving to center stage.

His conclusion was stark: “The five-century growth of an unabridged right of citizens to speak without controls may be endangered.”

Pool worried that shifting to new mediums of communication and information-sharing left a window — a period where the free expression implications of the new technological mediums were not obvious — for the state to establish a foothold of influence before the First Amendment could catch up.

His most important observation about these shifts was that they would occur without society realizing what was happening. Because these new technologies were understood as requiring technical oversight, “regulation seemed to be a technical necessity.” The emerging new centers of information-sharing would be brought under government control through what lawmakers and regulators believed to be technocratic, democratic, seemingly unobjectionable decisions. Only later, when those technologies were fully woven into our social fabric, would society realize the control over information it ceded to the government — perhaps to actors with malign intentions, like those Pool had carefully studied all those years earlier.

So what led Pool to his grave conclusion? Bitter experience.

Pool had long observed that regulators and courts often missed the expressive implications of new communications technologies by analogizing them to infrastructure rather than expressive tools, like the press. The telegraph, for example, was regulated like a kind of transportation network rather than a new vehicle for speech; its regulatory structure borrowed from the common-carrier model developed for railroads.

The sidelining of free expression was as much an outcome of social unawareness as government determination. Because early telegrams “carried so few words” at “such a high cost” and were primarily business-oriented, he said, censorship was not on anyone’s mind. In fact, “there were virtually no cases in which First Amendment rights were claimed or such precedents applied.”

Lawmakers likewise analogized the telephone to the telegraph and imposed a similar legal framework. Broadcasting followed the same pattern, ushering in nine decades (and counting) of direct government censorship of the content spread over the airwaves, showing the consequences of failing to recognize the full breadth of First Amendment protections for new expressive technologies.

And with the world of computers and the internet that was materializing at the time of Pool’s warning, heeding his message by extending First Amendment principles to those new technologies has been a long and ongoing fight.

Now, as legislative attention turns towards the next evolution in expressive technology, artificial intelligence, Pool’s warning is of paramount importance.

A Look Forward

In this series, we’ll dive deep into the First Amendment questions raised by artificial intelligence. The idea that text seemingly spun from the machinations of an incomprehensible black box enjoys protection similar to the words out of your mouth can be bewildering. A federal judge in Florida noted as much in a case for which FIRE filed an amicus curiae brief, the judge struggling with the notion that “words strung together by an LLM are speech.”

We’ll argue the technology is less inscrutable than it may first appear, and that applying First Amendment principles to AI-generated expression is both a natural extension of those principles and a necessary step to prevent those principles’ weakening in other applications, including in areas like newspapers and video games.

Opinions will differ on some points, like whose speech is implicated by restrictions on AI tools and where liability should be assigned. But, heeding Pool’s message, these debates must be grounded in the fact that AI systems are rapidly becoming a dominant way people seek information, ask questions, and develop ideas. From a First Amendment perspective, that is sacred terrain which we cannot afford to lose to government control.

The Supreme Court has made clear that technological change does not dilute that interest. As it reiterated in 2024 when talking about social media platforms, “whatever the challenges of applying the Constitution to ever-advancing technology, the basic principles of the First Amendment do not vary.”

Those words have a straightforward implication here. However novel AI-generated speech may seem, the First Amendment remains consistently concerned with state efforts to shape or control the flow of information and ideas, whatever form it takes. That’s our bedrock principle as we work through the particulars of this live issue.

It’s a principle that we believe is intuitive. Imagine if a Republican administration required AI outputs to align with conservative views on contested social issues, or if a Democratic administration required them to align with progressive views on those same issues. It would activate the same classically American aversion to censorship the respective sides would feel if the government was acting on a newspaper or a cable channel.

That instinct drives constitutional analysis, too. The Supreme Court has long refused to let labels muddle the essence in defining the scope of constitutional rights: “The Constitution deals with substance, not shadows.” That matters here. If the substance of an AI regulation controls what ideas an expressive technology may be used to generate, transmit, suppress, or prioritize, the First Amendment interest can not be overridden by framing AI as novel, technical, or as a mere ‘product.’ The medium may be novel. The constitutional problem is not.

Abandon that principle, and Pool’s warning becomes reality: a world where Americans will remain free to speak to their heart’s content in yesterday’s forums, but when it comes to the forums of tomorrow, expression will exist only by government permission.

Tune in next week where we’ll explain the basics of how the First Amendment applies to AI. See you then!

Illinois can tax income. It can tax profits. It can tax businesses. It can even impose generally applicable taxes that happen to reach content mediums like cable or newspapers.

Subscribe now

But the First Amendment strictly prohibits taxes that single out content the state doesn’t like. Unfortunately, that’s exactly what Illinois’ new state spending plan does, and it’s poised to soon be signed by Governor Pritzker.

Buried in the 1600-page budget, the relevant provision would charge the secretary of state with collecting a “social media platform fee.” Platforms covered by the proposal would pay the fee monthly, with the size of the bill set by “the number of Illinois users from whom the social media platform collects data within a month.”

Pinpointing where exactly the problems that doom this proposal begin is a tall order.

Doomed in the courts

The proposal’s biggest hurdle is the decades of case law that have squarely labeled this kind of tax as exactly what it is: a regulation of speech.

And social media sites are very much speech. As the Supreme Court stated in the recent landmark case Moody v. NetChoice, “To the extent that social-media platforms create expressive products, they receive the First Amendment’s protection.” That means the government can’t restrict or single out their content for regulation, and that includes singling them out for taxation.

That principle did not begin with social media, or even modern First Amendment law. In fact, it dates all the way back to England in the early 18th century, and in the American colonies with the resistance to the 1765 Stamp Act. That Act, which required publications such as pamphlets and newspapers to pay a tax and carry a special government stamp, sparked the first unified protests against British rule that eventually led to the Revolution.

In Grosjean v. American Press Co., the Supreme Court explained that English taxes on newspapers and advertisements like those introduced in the Stamp Act were not simply revenue measures. They were understood as efforts to suppress criticism of the government by making publication and circulation more difficult. These “taxes on knowledge” helped animate the First Amendment’s strong protections, and they influenced the outcome in cases like Grosjean.

In that 1936 case, the Court weighed a Louisiana tax on newspapers passed by allies of Louisiana Senator Huey Long. The plaintiffs argued that Louisiana’s 2% tax on newspapers with weekly circulations above 20,000 copies targeted the papers most critical of Long — papers he and his allies would like to see less circulated.

That connection between the financial burden of the tax and the potential circulation of the newspaper is exactly why the Court said such taxes “operate as a restraint” on speech. “First, [the tax’s] effect is to curtail the amount of revenue realized from advertising” and “second, its direct tendency is to restrict circulation.”

The connection the Court draws in Grosjean between the financial burden and the circulation of speech is relevant to the Illinois bill in more ways than one. On top of the platform fee described above, the Illinois budget proposal also imposes a 10% digital targeted-advertising tax on receipts from targeted ads provided in the state, a tax which expressly includes “advertising on social media.” Following the logic of the Grosjean Court, it’s another way the Illinois proposal discourages the speech it disfavors.

The holding that taxes on speech are speech regulation doesn’t end with Grosjean.

In 1983’s Minneapolis Star v. Minnesota Commissioner of Revenue, Minnesota imposed a special use tax on paper and ink used by publications, with an exemption that meant only a handful of publishers paid significant amounts. Unlike in Grosjean, there was “no indication . . . of any impermissible or censorial motive” behind Minnesota’s tax. But that did not save it.

That’s because the First Amendment problem was structural: Minnesota had “singled out the press for special treatment.” In burdening a specific group of speakers, the Court held the law was presumptively unconstitutional. The Court made clear that “illicit legislative intent” is not necessary for there to be a First Amendment violation.

While the Illinois proposal similarly singles out a group of speakers — social media platforms — it also categorizes them based on their content: the burden applies only to platforms where users “create, share, and view user-generated content.” Arkansas Writers’ Project, Inc. v. Ragland is instructive. There, the Court struck down an Arkansas sales-tax scheme that exempted newspapers and “religious, professional, trade and sports journals and/or publications printed and published within this State,” but not general-interest magazines. The Court characterized that content-based burden as “even more disturbing” than the speaker-based burden in Minneapolis Star. Again, it was presumptively unconstitutional.

This line of cases was further backed up for the internet age by the federal Internet Tax Freedom Act, which prohibits discriminatory taxes on electronic commerce. By “targeting large online platforms without a comparable tax on offline media or communication services,” the Illinois Policy Institute concludes the social media tax proposal could be doomed by federal law before even getting around to the First Amendment arguments, not to mention its other problems.

Doomed in the details

The proposal’s constitutional problems are reason enough to reject it. But even setting those aside, its details are riddled with flaws that doom the implementation of the proposal to a mess of confusion and unnecessary burdens.

First, its sloppy definition of which platforms are covered by the proposal would leave a plethora of websites outside the social media umbrella as collateral damage.

Per the bill’s language, “[c]overed platforms” are platforms where users “create, share, and view user-generated content” that “can be viewed by other users of the medium” and “primarily serves as a medium for users to interact with content generated by other users of the medium.”

That casts a wide net.

While the definition would surely include X and Facebook, the list of sites on which users interact with other users’ content is long. That could include publishing sites from Substack to FanFiction.Net, messaging services, review sites, and even crowdsourced fan wikis.

Often, statutory definitions like this include a range of carve-outs to help guide our understanding of their language. Instead, Illinois leaves us with a single explicit exemption: “‘Social media platform’ does not include a not-for-profit organization.”

The proposal’s vague guidance could have straightforward consequences. Content-sharing sites with large audiences but limited revenue streams often run on lean budgets, unlike the sprawling social media giants this bill seems designed to target. Imgur, for example, had at one point 30 million monthly users but two employees.

The lack of clarity around whether a small team like that is responsible for a large monthly fee could be an existential question. This could chill the inclusion of collaborative features that platforms might fear will land them a visit from the tax collector — sorry, the secretary of state.

Platforms are also left to their imagination to discern what exactly counts as a “user” for purposes of calculating the tax. Is a user an account? A unique individual who might have multiple accounts? What about simple lurkers of the site with no account? The ultimate answer could make a big difference in the monthly fee platforms will plan to face.

We’re not surprised to see a proposal like this filled with ambiguities and holes. As states have attempted to turn around the losing streak social media regulation has faced in court, we’ve seen a number of half-baked schemes to find some kind of creative workaround. States have tried to avoid scrutiny by casting their speech restriction as regulating “design choices,” by targeting platforms through the app store, and by labeling their target “conduct” rather than “speech.”

Fortunately for free speech, the courts aren’t impressed. An attempt to punish speech is an attempt to punish speech, any way it’s structured.

This essay was originally published by RealClearPolitics on May 8, 2026.

The dawn of the AI era has sparked a wide range of reactions, from exhilaration over the technology’s capabilities to deep distress.

Such responses to a new communicative tool are nothing new, and indeed, AI presents new and unique challenges that will require deep thought and sensitivity.

Subscribe now

But a heavy-handed congressional response that erodes longstanding American freedoms isn’t the answer. The Senate Judiciary Committee’s markup and passage last week of SB 3062, the GUARD Act, shows the substantial risk that Congress’s “do something” energy poses to the speech rights of everyone.

The bill regulates AI chatbots — especially so-called “AI companion” systems — through access limits, design mandates, and disclosure requirements, backed by civil and criminal penalties of up to $100,000 per violation. If enacted, it puts the federal officials squarely in the position of deciding how this technology is built and used, limiting engagement with information and compelling speech along the way.

Growing calls for a federal solution, including from the White House, to fix the fragmented landscape of state regulations reflect a clear political appetite for legislative action. And a single national standard has obvious appeal for an industry seeking consistency across jurisdictions. But consistency isn’t the same as constitutionality. If federal proposals like the GUARD Act replicate the speech restrictions found in state laws, they just hardwire those problems into federal law.

Take the bill’s age verification requirements. The GUARD Act forces Americans to create accounts and prove their age, with minors barred from some “AI companion” systems. Existing accounts are frozen until verified, and companies are required to recheck users’ ages periodically.

Age-verification mandates like this one force individuals to disclose their identity to seek answers and thus give up anonymity, a right the Supreme Court has repeatedly recognized as central to free expression. Faced with mandatory identity disclosure, many think twice before asking sensitive questions. Would someone trapped in an abusive relationship be more, or less, willing to seek advice from a chatbot if they had to surrender their privacy? Or how about the employee who is consistently harassed at work but is worried about asking for advice? There’s a reason that the Federalist Papers were written under the pseudonym “Publius” — even public debate sometimes requires distance from the speaker’s identity. That protection still matters today, allowing people to seek information, test ideas, and ask sensitive questions without fear of legally required exposure.

Then there are rules about content. The bill makes it unlawful to design, deploy, or make available chatbots that, in the government’s view, “encourage” or “promote” certain categories of constitutionally protected speech. Who do we want to be in charge of determining that? Those restrictions violate the First Amendment by regulating the protected editorial decisions of developers and by infringing on individuals’ rights to create and receive lawful expression.

Proposals like the GUARD Act dictate how chatbots respond and intrude on editorial judgment by putting Congress’s thumb on the scale of what’s acceptable speech. This means control over who can speak, what can be said, and how ideas are expressed.

Those choices shape the substance of speech and risk reducing a chorus of voices to a single, government-shaped note. Grok is loosely modeled on The Hitchhiker’s Guide to the Galaxy. Claude operates under its own internal “constitution.” Standardizing those philosophical differences flattens distinct approaches, and when fewer questions are asked, fewer answers follow.

Finally, disclaimer mandates can cross constitutional lines by compelling speech. The GUARD Act requires chatbots to deliver federally imposed messages in every interaction. While informing users, its application in every circumstance alters the content and flow of communication itself, overriding both user and developer choices with what the officials want the public to see.

All of this points to a deeper reality that AI systems cannot perfectly predict or control every output. That’s not a defect. It’s a core feature of how these models generate responses from probabilistic patterns. Developers will be forced to filter even more speech than the bill directly targets to ensure the offending content is not generated. Combined with the GUARD Act’s vague and sweeping restrictions, the result is blunter tools that sand down the rough edges of debate and offer less of what makes them useful in the first place.

Treating chatbots as expressive tools keeps the focus on the people, not the machine. Many who’ve used them — which is now well more than half of Americans — know about its potential. It lets people test arguments, explore unfamiliar ideas, and tackle everyday challenges.

Yet artificial intelligence, particularly chatbots, has become Washington’s latest political punching bag. Accusations of manipulation and harm are driving a slew of legislative proposals to censor this emerging technology. The GUARD Act isn’t alone. The recently introduced CHATBOT Act presents many of the same threats.

The same impulse to move quickly in Congress is playing out nationwide, with proposals in states like Minnesota, Florida, and Washington targeting chatbots through access restrictions, disclosure mandates, and content-related rules. Florida Gov. Ron DeSantis is pushing an AI package that effectively requires everyone — adults and minors alike — to identify themselves before using these systems.

But the Constitution doesn’t permit any government to address concerns about AI by broadly restricting protected expression. The First Amendment demands solutions that target illegal conduct without burdening the exchange of ideas.

Today, AI sits squarely in the hot seat. Lawmakers and cultural commentators alike argue this moment is different — that the scale, speed, and influence of modern networks justify new forms of oversight and restriction.

But history suggests a recurring pattern: efforts to curb perceived harms often expand beyond their original aims, reshaping the boundaries of speech and expression in profound and lasting ways. As we confront this latest wave of tech debates, the question is not whether change is coming — it’s here. The question is which lessons we should learn from the past.

On March 24, 2026, at an intimate venue in Manhattan, we brought together some of the voices examining these recurring cycles. Nico Perrino, FIRE’s executive vice president and host of So to Speak: The Free Speech Podcast, places today’s debates in the longer civil-libertarian tradition of resisting censorship. Corbin Barthold, internet policy counsel at TechFreedom and host of the Tech Policy Podcast, brings a litigator’s skepticism to the latest proposals to regulate online platforms as common carriers. Christopher J. Ferguson, professor of psychology at Stetson University and co-author of Moral Combat: Why the War on Violent Video Games Is Wrong, has spent years studying the evidence behind recurring fears about violent media and new technologies. And guiding the discussion is Kmele Foster, co-host of the Fifth Column Podcast and editor-at-large at Tangle News, whose work is defined by intellectual curiosity and a rigorous commitment to open inquiry.

• Nico Perrino (X), FIRE’s executive vice president and host of FIRE’s So to Speak: The Free Speech Podcast

• Corbin Barthold (X), internet policy counsel at TechFreedom and host of the Tech Policy Podcast

• Christopher J. Ferguson (X), professor of psychology at Stetson University, author of How Madness Shaped History, and a licensed clinical psychologist

• Kmele Foster (X), editor-at-large at Tangle News, and partner at Big Think

Recent reports suggest the Trump administration is now considering new oversight for advanced AI models like ChatGPT, Claude, and Gemini. Few details have been finalized, but officials are reportedly discussing an executive order to create a government–industry working group. Another idea under consideration is a process for reviewing models before or around their release.

Subscribe now

As these talks move forward, they risk setting a troubling precedent for free expression.

Some AI companies such as xAI have already agreed to provide an early release of their models. One report announced that a group has already been formed within the Department of Commerce, which will review these models — sometimes testing them with fewer safety limits — to see what risks they might pose, especially for cybersecurity and national security.

This isn’t entirely new as OpenAI and Anthropic have done something similar before. But now the Trump administration is considering making this kind of review much more formal.

AI is an expressive tool — one that people use to learn, ask questions, and engage with ideas. The design of an AI system also reflects a series of choices about what information to prioritize, how to reason, what boundaries to draw, and how to respond. Those choices embed values and assumptions about knowledge, truth, and human interaction.

People who build and use AI tools do not shed their constitutional rights to freedom of expression at the prompt window. That includes the right to speak without being pressured by the government to first seek approval or give them a look under the hood. What starts as “just a review” can quickly become pressure to change what tools the public is allowed to have and what information users are allowed to see. Informal oversight has a way of turning into coercion.

Imagine an AI company being told its approval depends on how its model handles controversial issues — whether it reflects the government’s preferred stance on tariffs, energy or climate policy, or how it discusses election integrity ahead of an upcoming vote. Even without an explicit mandate, that kind of signal would pressure developers to shape outputs to satisfy officials rather than reflect independent judgment.

We’ve seen this before. When Biden administration officials approached social media platforms and repeatedly pressured them to remove or downgrade COVID-related content, it meant the public discourse about the most important issue of the day was being covertly shaped by government officials. In the current Administration, we’ve observed how the Federal Communications Commission’s authority to regulate the broadcast spectrum — which sounds fairly technical and benign in theory — has been weaponized to pressure news outlets to reshape their coverage at the government’s behest.

In those cases, pressure largely operated after content was posted. This new arrangement is more dangerous because it moves that pressure upstream, before an AI model is released to the public. Giving the government a role in reviewing speech or expressive tools at that stage creates a point of leverage it can use to ensure certain expression never sees the light of day. In First Amendment terms, this amounts to prior restraint — blocking speech before it’s communicated, rather than punishing it afterward — if the government can give thumbs-up-or-down approval. That’s why they’re presumed unconstitutional by courts.

Officials often invoke national security to justify expanding their reach. But national security isn’t a blank check. The public should be wary of arrangements that give government officials of either party a foothold over what AI systems generate.

Once the government assumes a gatekeeping role over emerging forms of expression, the line between oversight and censorship gets blurry real fast.

Artificial intelligence is rapidly becoming one of the world’s primary engines of knowledge creation. And as this meaning-making system comes online, so too does the inevitable desire for government control. How the First Amendment and other legal principles apply to AI will determine whether the future operating system for the planet acts as a tool for inquiry and expansion of human knowledge and capacity — or for compliance and top-down control. History is blunt on this point: once government acquires durable power over speech and expression, that power only ever expands — and the lure of censorship becomes virtually impossible to resist.

Subscribe now

This discussion — from Jan. 22, 2026 — brings together the individuals tracking and resisting that expansion. Dean Ball — technologist, AI governance analyst, and former White House science and technology adviser — warns that “responsible AI” rhetoric often serves as an on-ramp to permanent regulatory control. Ari Cohn, First Amendment attorney and tech policy expert, brings a free speech lawyer’s clarity to where government proposals cross constitutional lines. Matt Perault, head of AI policy at Andreessen Horowitz, examines how regulatory choices shape the incentives that drive innovation. And guiding the discussion is Kmele Foster of The Fifth Column podcast, whose work is defined by a first-principles approach and an inimitable commitment to curiosity.

Panelists:

• Dean W. Ball (X), senior fellow at the Foundation for American Innovation, former senior AI policy advisor at the White House OSTP, and principal staff drafter of “America’s AI Action Plan”

• Matt Perault (X), AI policy head at Andreessen Horowitz (a16z), and former director of the Center on Technology Policy at University of UNC Chapel Hill. Read a16z’s AI policy work on Substack

• Ari Cohn (X), a nationally recognized attorney and expert on the First Amendment, defamation, and Section 230

• Kmele Foster (X), editor-at-large at Tangle News and partner at Big Think

The encrypted messaging app Signal is back in the news — and this time, people are asking: Will using it get me arrested?

The short answer is probably not. If your speech would be protected if you were talking to a friend at a park or over the phone, your speech will still be protected when you have that conversation on Signal. And if not, you’re not any more likely to get arrested by using Signal than you were before you started using it. In other words, if you’re doing something unlawful, adding Signal to the mix probably doesn’t appreciably increase the already existing risk of getting arrested for breaking the law.

Let’s look at why that’s so.

Why are people suddenly worried about Signal?

In January, independent conservative journalist Cam Higby posted on X that he had “infiltrated organizational Signal groups all around Minneapolis with the sole intention of tracking down federal agents and impeding/assaulting/and obstructing them.” In his thread, Higby shared screenshots showing how some anti-ICE groups were organized along with the screen names of some participants, which led others to speculate on connections between organizers and Minnesota political figures.

Two days later, on the Jan. 26 episode of the Benny Johnson podcast, Johnson asked FBI Director Kash Patel about Higby’s reporting. Patel replied that the FBI had opened an investigation into Signal group chats, saying (emphasis added):

As soon as Higby put that post out, I opened an investigation on it. Just like any other case . . . We immediately opened up that investigation because that sort of Signal chat being coordinated with individuals, not just locally in Minnesota, but maybe even around the country . . . if that leads to a break in the federal statute or a violation of some law, then we are going to arrest people.

The significance a listener attributes to this quote is going to depend entirely on what they think of the actions of anti-ICE protestors in Minnesota. If the listener thinks the protestors are intending to harm the agents or interfere with federal law enforcement, then the FBI investigating an alleged coordinated effort to commit crimes is probably unremarkable. That’s what the FBI is supposed to do, at that level of abstraction.

On the other hand, if the listener thinks the protestors are engaged in protected First Amendment activity, then the FBI opening an investigation into Signal chats with the promise of arresting people if a crime is committed sounds a lot like viewpoint discrimination. Higby’s thread did not itself show evidence of illegal activity — just people sharing information about law enforcement activity, which is generally protected speech. (FIRE is suing the government on behalf of individuals who created a Facebook group and an app, respectively, that share video of ICE agents doing their jobs.) Many have used that information to protest or film ICE agents, which the First Amendment also protects. Nothing sounds more authoritarian, and less American, than the government investigating people for not liking the government.

But this piece is focused on a secondary effect of Patel’s comments: the fear that downloading Signal is going to put you in the FBI’s crosshairs.

How secure is Signal compared to other messaging apps?

What is it about Signal that has everyone’s attention? To understand why it’s popular among protesters and the Pentagon alike, let’s look at how messaging apps normally work — and how law enforcement might normally use data from those apps to investigate crimes. To do that, we need to think about two things: messages (that is, the actual content of what you’re sending) and the metadata (which is some combination of things like the time sent, the time read, location, device, phone number, network carrier, and so on).

In unsecured text messaging, like SMS or Facebook messages, your messages are sent to a service provider, where they’re stored along with the metadata. Cell phone providers store SMS data, although they generally don’t store it for very long. But in 2022, a detective subpoenaed Facebook messages to investigate crimes related to a suspected unlawful abortion by a 17-year-old. Those texts revealed the teen’s mother had been coaching her daughter on how to take abortion pills, leading to the mother being charged as well.

Next are messaging apps where the messages are end-to-end encrypted — meaning they can only be read by the devices on either end of the conversation — but the metadata is saved. One example of this is WhatsApp. This means the service provider doesn’t know what you’re saying, but it knows who you said it to, and when, and how often, and sometimes, where you were when you said it, and what phone you were using. For law enforcement, that’s often more than enough. In 2021, former Treasury official Natalie Edwards was sentenced to six months in prison for leaking documents to the press. When investigating Edwards, the FBI got a court order to access the metadata from her WhatsApp conversations. They couldn’t see what she was talking about, but they could see she was having frequent conversations with a reporter, and how those times lined up with stories containing leaked data.

(Side note: Apple’s iMessage is a special case because the messages are end-to-end encrypted, but the default behavior of iCloud is to store the encryption keys in the cloud for data recovery. That’s a setting you can change, if you want. But the default behavior is very much the digital equivalent of locking up your house and sliding the key under the mat.)

And then there’s Signal, which end-to-end encrypts its messages and the only metadata it collects is account-creation time and last-connection time. When the FBI came knocking in 2021, that’s all Signal had to turn over. That’s probably why anti-ICE protesters are using it, why the government has used it, and why tens of millions of other people have, too. It’s also open-source, so it can be (and has been) reviewed for any backdoors, which are undocumented methods of bypassing access control measures (something the government has called on companies to add).

Does that mean Signal is perfect? Not exactly, because it exists in the same imperfect world the rest of us do. For example, if you’re using your cell phone to open Signal, there will still be cell tower data about the time you were using your phone, and your cell provider might have metadata that your device communicated with Signal’s servers. Using a VPN could help with the latter, but then it’s a question of how much you trust your VPN. In theory, if you and the person you’re communicating with already have your metadata being captured by these companies, maybe just opening the app around the same time, repeatedly, could be evidence. And, of course, humans are always vulnerable to opsec failures.

But it’s fair to say that trying to reconstruct a crowd of human connections from fragments of metadata scattered across multiple companies is nontrivial. Even if the FBI could get large swaths of metadata — and maybe it can, if it buys it on the open market — it would probably need an immense neural network to try to piece together these fragments into coherent relationships. And where would the FBI get an artificial intelligence to use in that kind of mass surveillance, anyway?

FIRE is working on these issues. In an amicus brief filed last week in support of Anthropic’s lawsuit against the government, we argued:

For example, an agency could use an LLM to infer an individual’s association with a particular mosque based upon frequent visits to the mosque’s website, engagement with the mosque’s social media posts, and their cell phone’s physical proximity to the mosque during religious services. [...] It is easy to conceive how an agency, a government employee with improper intent, or a malicious third party that finds a vulnerability, could exploit these capabilities to monitor public discourse, preemptively squelch dissent, or cause myriad other harms.

A few days earlier, we joined an amicus brief in Chatrie v. U.S., arguing that geofence warrants (that is, warrants that compel service providers to hand over location data for every device near a location during a specific time window) violate First Amendment rights. Protestors, religious congregants, journalists meeting confidential sources, and innocent people who might happen to be nearby all have their rights chilled when the government engages in these digital dragnets.

Does the First Amendment protect the use of encryption like Signal?

Ultimately, Signal is encrypting your messages, and there’s good reason to suspect both the First Amendment’s protection of speech and the general right to privacy protect your messages from government intrusion. There are a few sources of law we can look to for those principles.

There should be books written about Bernstein v. U.S., and if there are, well, there should be more of them. After World War II, the U.S. government broadly regulated the export of cryptography (that is, a key that controlled the reversible transformation of a message) but did not regulate one-way hashing (a mathematical formula that would turn one string into another and could not be reversed). But are those things really different? Isn’t it all a form of speech?

In the 1990s, then-mathematics doctoral student Daniel Bernstein submitted for government review a program called Snuffle. I’ve been told I lose people when I try to explain how it worked, so I’m going to abstract this into really simple terms: Snuffle showed how it was possible to make working encryption in a way that didn’t meet the government’s technical definition for export controls. (It didn’t even actually do it. The code was missing essential pieces. It just showed how to do it.)

The government balked, and Bernstein sued. In its opinion, the Ninth Circuit panel said that Snuffle was entitled to First Amendment protection:

In this increasingly electronic age, we are all required in our everyday lives to rely on modern technology to communicate with one another . . . Government efforts to control encryption thus may well implicate not only the First Amendment rights of cryptographers intent on pushing the boundaries of their science, but also the constitutional rights of each of us as potential recipients of encryption’s bounty.  

The Ninth Circuit agreed to rehear the case before the full court, vacating the earlier panel decision. But before the case could be reheard, the government loosened the export rules and subsequently dropped the appeal. That procedural quirk rendered the panel’s decision void, but its wisdom still holds. In the 2000 case Junger v. Daley, the Sixth Circuit rejected the argument that computer source code isn’t really speech, noting that “though unintelligible to many, [it] is the preferred method of communication among computer programmers.”

Encryption aside, nothing about the use of Signal to organize groups that monitor federal activity appears to be unlawful. We have a right to organize for political action. We also have the right to monitor law enforcement with phones. And the Fourth Amendment protects us from warrantless collection of our location information or warrantless searches of our cell phones.

Can Signal keep you from getting caught if you commit a crime?

Signal isn’t a magic get-out-of-prosecution-free app for the same reason that no system is ever going to do that — because once you let humans in, the system becomes imperfect. We started down this road not because of a subpoena, but because someone just joined the groups planning the activity and then told everyone what was going on.

So that’s the hot takeaway: Don’t commit crimes, because you’ll get caught. That’s got nothing to do with Signal though. Signal’s fine.

Across the United States, teenagers freely express themselves online. But that freedom is rapidly being restricted, and make no mistake: this doesn’t just end with teens. What is often portrayed as a youth mental health issue is really a battle for everyone’s online speech rights.

Subscribe now

Worldwide, governments are moving to restrict minors’ access to social media. In Europe, France is seeking to ban children under 15 from using social media without parental consent, the United Kingdom is pursuing expanded regulatory authority over youth online access and platform design, and Spain has proposed criminal liability for tech executives who fail to remove harmful content quickly enough. In Asia, China has tightened limits on screen time and platform use for minors while Malaysia is considering a nationwide ban on social media for users under 16. And in Australia, the government has approved sweeping youth access restrictions.

The United States has shown signs of taking a similar course. Congress is advancing legislation like the Kids Online Safety Act (now combined with other bills in the House under the title “KIDS Act”) and multiple states have enacted or attempted age-based social media restrictions despite ongoing First Amendment challenges. The policy momentum is unmistakable — and troubling.

The concerns driving it are, to some extent, real. Nearly half of American teens say social media negatively affects people their age. Public health authorities have linked heavy social media use to anxiety, depression, sleep disruption, and increased exposure to cyberbullying. Research from the World Health Organization has documented declining adolescent well-being associated with excessive screen time and digital pressure. At the same time, other studies have found beneficial effects of social media use, particularly for vulnerable youth seeking community and support, and researchers continue to debate whether many of the observed harms have been causally proven.

As a 20-year-old who grew up during the rise of Instagram, Snapchat, and TikTok, I recognize those harms. I remember when social media shifted from something you checked occasionally to something that shaped your social world. Group chats determined inclusion. Posts felt permanent. Comparison was constant. I have felt the pressure of likes and the anxiety of visibility. I have watched peers struggle with online harassment and digital burnout.

Today’s youth are experiencing troubling mental health trends. But even if social media does play a role, when authorities claim a noble and urgent purpose in regulating how people can express themselves, these proposed solutions demand scrutiny.

Age-based social media laws do not simply reduce screen time. Most rely on age verification systems that require identity verification in practice. That can mean uploading government-issued identification, biometric scans, or other sensitive personal data just to create an account. What is framed as child protection can quickly become a structural shift in how speech is accessed online, and how many invasive barriers the government requires private companies to place between their users and their platforms.

From a civil liberties perspective, that shift is significant. The Supreme Court has repeatedly recognized the importance of anonymous speech in American tradition, such as in cases like Talley v. California, and McIntyre v. Ohio Elections Commission. Anonymity protects political dissidents, whistleblowers, vulnerable communities, and young people exploring their identities. If participation in digital discourse increasingly requires identity verification, anonymity weakens and the chilling effect on lawful speech grows.

Strict liability regimes create additional risk. If platforms face legal penalties for failing to remove harmful content quickly enough, which is a tool regulators in Europe especially rely on, they will inevitably err on the side of removing more speech. Automated moderation systems cannot perfectly distinguish between offensive but lawful speech, unprotected speech, and speech that is legal but nevertheless targeted by lawmakers as subjectively “harmful.” Discussions about politics, religion, gender, or social justice may be flagged and suppressed simply because platforms cannot afford the regulatory risk.

Supporters argue these laws target only minors. But the technical infrastructure required to verify age often requires verifying everyone. Platforms cannot easily build parallel systems for adults and children without expanding data collection across the board. That means more identification checks, more stored data, and greater vulnerability to breaches or misuse.

There are better ways to address youth mental health without reshaping the architecture of free expression. Platforms can voluntarily provide stronger parental control tools and resources that help families manage online experiences. Schools can invest in digital literacy education that teaches healthy online engagement. Efforts to address youth well-being should focus on empowering users and families rather than imposing government mandates that reshape how speech is accessed online.

The real question isn’t just teen safety, but this: Is the United States willing to normalize a permission-based internet?

My generation understands the downsides of social media. We have experienced comparison culture and digital pressure. But we have also used these platforms to organize protests, build communities, share stories, and participate in civic life. Social media is not merely entertainment. It is a modern public square, and when we understand it in those terms it’s clear why we cannot tolerate government intervention that censors speech and decimates anonymity.

Protecting teens is essential. Preserving free expression is essential, too. The vast majority of even “harmful” speech on social media is still protected speech that minors have a First Amendment right to access. If access to the public square increasingly depends on proving who you are before you speak, we may address one problem while creating another: a more monitored, less anonymous, and ultimately less free digital environment for everyone.

Lawmakers should approach youth social media reform with care. The goal should be to reduce harm without blocking minors from content they are entitled to, and without infringing on everyone’s rights by conditioning online speech and giving up anonymity.

In April of 2023, 14-year-old Sewell Setzer III became a user of Character.AI, a platform founded by two former employees of Google that hosts user-created interactive chatbots inspired by popular fictional properties. Going by “Ageon,” “Daenero,” and other names, Setzer began an intimate correspondence with a Game of Thrones-inspired “Daenerys Targaryen” chatbot. Less than a year later, he had killed himself. To Setzer’s family, his final exchange with Daenerys pointed to Character.AI as the culprit.

The leading subject of a New York Times feature just last week, Sewell’s parents’ wrongful death suit — filed in October 2024 — initiated what has become a growing wave of lawsuits seeking millions of dollars in damages from chatbot-based platforms. Each new plaintiff points to the last as evidence of a causal link between chatbots and the deaths of young users — and we’ve gradually seen the allegations take a turn from encouraging suicide to concocting elaborate plots.

Subscribe to Notice and Takedown

In August 2025, OpenAI found itself in the crosshairs with a lawsuit alleging 16-year-old Adam Raine’s suicide had been assisted and inspired by his interactions with ChatGPT. That same month, Stein-Erik Soelberg killed his mother and then himself, with his estate alleging that ChatGPT had convinced him he was the target of a high-level conspiracy. In November, the death of Austin Gordon by self-inflicted gunshot wound marked one of seven more high-profile lawsuits against OpenAI. His mother’s complaint, filed last month, alleged ChatGPT “created a fictional world and relationship that felt more real to Austin than anything he had ever known.”

And then somehow, things got even stranger. In the last two weeks, two lawsuits were filed:

One alleged ChatGPT essentially goaded a woman into torching a settlement agreement between her and the life insurance company that filed the suit, firing her lawyers, and engaging in a flurry of frivolous legal filings against the company. Another, filed against Google, alleges that their chatbot service Gemini had “trapped” 36-year-old Jonathan Gavalas in a “collapsing reality,” which involved coaching him through “missions” involving violence against the public and eventually his suicide.

Settling for less

The plaintiffs in these cases seem to be finding some success. Last month, Character.AI agreed to settle the Setzer-Game of Thrones case (styled Garcia v. Character Technologies) along with three other similar lawsuits filed in September.

They’ve been supported by political headwinds. Each of the complaints filed made a point to emphasize a letter by 54 state attorneys general warning of a “race against time” to “protect the children of our country from the dangers of AI,” insisting that the “walls of the city have already been breached.”

For anyone familiar with the history of civil liberties disputes, this rhetoric is instantly recognizable. Phrases like “race against time” and “protect the children” are the lingua franca of government restriction of novel expressive technology. The phrase “protect the children” gives you the license to restrict speech and the novelty of the technology provides the high stakes — if nothing is done before the technology develops past a certain point, it will presumably be too late. We have to restrict speech and we have to restrict it now.

The emotional appeal is strong. The defendant AI companies are inclined to avoid trial for a reason.

Is anyone raising First Amendment concerns?

You bet.

FIRE intervened early in Garcia after the court denied Character.AI’s motion to dismiss. In that order, the judge questioned why “words strung together by an LLM are speech.” A federal judge musing that stringing words together might not be speech because of who did the stringing together would depart from a long line of court cases holding that pure speech is protected speech agnostic to the identity of the speaker.

The order had worrying First Amendment implications for expressive technology if left to stand at the final conclusion of the case. FIRE accordingly filed a friend-of-the-court brief urging prompt appellate review of this holding, outlining all the reasons that statement — and the logic underpinning it — ran afoul of the First Amendment.

For the purposes of this blog, we’ll be evaluating the claims of these lawsuits in a bigger context. As case after case piles up, it is tempting — and quite human — to let the recurrence of tragedy take on a similar role as authoritative data in how we process the phenomenon, and importantly, assign blame. The public and the courts have confronted this temptation before.

We’ve rolled these dice before

There is a long line of entertainment-related torts and moral panics that have besieged free expression over the years, placing blame for violent acts on everything from Grand Theft Auto to the Slender Man lore. Each and every panic, taken to its logical conclusion, would have shrunk the universe of allowable expression in ways that would reverberate long past the point where clarity makes society’s past worries seem a little silly in retrospect.

No recent panic quite matches the intensity and the surreality of the current moment like the Dungeons and Dragons scare of the 1980’s. During a roughly five year period in the 1980’s there were 28 cases of adolescents who played Dungeons & Dragons and later committed murder or suicide.

There was the case of 17 year-old player James Dallas Egbert III, whose disappearance into nearby woods inspired speculation from the press that he had lost the ability to distinguish between himself and the game character he role-played. There was also 16-year-old Irving Pulling, whose death inspired his mother to start the public advocacy group “Bothered About Dungeons & Dragons” (BADD).

The media ran with it. BADD was featured in a 1985 60 Minutes segment that will help give readers a sense of just how strong the panic was, marginalizing experts with arguments from emotion. “The families who have suffered the loss of a loved one would disagree,” the narrator says, as the muted objections of a skeptical clinical expert play in the background. “If you found 12 kids in murder-suicide cases with one common factor,” he presses, “wouldn’t you question it?”

With the clarity of hindsight, the math finger-paints a pretty silly picture. “By 1984, 3 million teenagers were playing Dungeons & Dragons in the United States and the baseline suicide rate of adolescents overall would have been about 360 suicides each year,” University of Virginia professor of pathology James Zimring has pointed out. “So, when you look at the bottom of the fraction, at the denominator, Dungeons & Dragons was, if anything, protective. It had the opposite effect.”

We shouldn’t have to wait for the chatbot panic to be in the rearview mirror to do the same math with the 13-18 million teenagers and 130 million adults using ChatGPT and other AI chatbots. When you consider the small number of (emotionally potent) cases, it begins to look like maybe AI is causing psychosis — just not in the way people think.

Exploding books and dangerous ideas

It’s not just “standard” First Amendment law that these lawsuits get wrong. In an effort to get as far away from speech as possible, plaintiffs’ lawyers have gone with products liability law. After all, who could argue with the idea that a company has an obligation to design safe products, right?

But when you drill down into it, they aren’t really talking about “products” at all.

The Garcia case alleged, for example, that Character.AI designed products that caused users like Sewell to “conflate reality and fiction.” That should sound awfully familiar; it’s basically the same accusation grieving mother Sheila Watters made in 1989 against Dungeons & Dragons maker TSR.

As the court’s decision in Watters v. TSR, dismissing the suit describes, she “cast her son as a ‘devoted’ player of Dungeons & Dragons, who became totally absorbed by and consumed with the game to the point that he was incapable of separating the fantasies played out in the game from reality.” According to her suit, this made the product (i.e., the game) “unsafe” and TSR should pay.

But the Watters Court rejected this theory of liability — the same theory underlying most if not all of the chatbot lawsuits.

The Sixth Circuit, upholding the district court’s dismissal, observed that the harm originated not from the tangible properties (or even rules) of the game, but rather from the ideas expressed through its storyline — and that meant the case wasn’t really about a defective “product.” A court examining claims that violent video games caused the Columbine shooting reached the same conclusion: “There is no allegation that anyone was injured while Harris and Klebold actually played the video games . . . The actual use of the video games, then did not result in any injury . . . So, any alleged defect stems from the intangible thoughts, ideas and messages contained within.”

That’s an important distinction — product liability is generally imposed (often without requiring any fault, referred to as “strict liability”) on tangible “products” (think brakes, tires, dishwashers, etc.) with inherent and unreasonable dangers that are hidden to consumers, or for which there is a safer design — putting the manufacturer in the best position to prevent harm. In other words, the physical thing hurts you physically.

Imagine that you purchase a book. If the book’s binding explodes when you open it, you’ve got a product liability claim. The physical book, regardless of what its pages say, exploded in your hands — and there’s no harm to free expression by saying you can’t sell a book that doubles as an IED.

But suppose you were harmed because you did something stupid after reading ideas in a book. You might be able to see how imposing liability for “dangerous” ideas would set us down a dark path; every author and publisher would have to make sure that the ideas they put out in the world couldn’t possibly be interpreted or used to some harmful end. If you’ve ever met other human beings, you already know that the list of such ideas is … quite short.

And that’s exactly what drove the outcome in Watters. The district court noted that “the theories of liability sought to be imposed . . . would have a devastatingly broad chilling effect on expression of all forms . . . The First Amendment prohibits imposition of liability . . . based on the content of the game.” The appellate court saw a similar unavoidable impact of allowing for such liability: “The only practicable way of ensuring that the game could never reach a ‘mentally fragile’ individual would be to refrain from selling it at all.”

Tale as old as time, song as old as rhyme

This understanding has been applied across mediums of content and entertainment. In the cases of McCollum v. CBS, Inc. and Vance v. Judas Priest, the musical artists Ozzy Osbourne and Judas Priest were sued over the idea their music encouraged the suicide of two young men (attempted suicide in the case of Vance). Like Watters and like the recent chatbot cases, the plaintiffs were families of the young men.

Their lawsuits were unsuccessful. The court in McCollum echoed the Watters court concerns about liability chilling the expression of creators, making clear “such a burden would quickly have the effect of reducing and limiting artistic expression to only the broadest standard of taste and acceptance.” They accordingly noted that in the history of attempts to assign tort liability for electronic media inciting unlawful conduct, “all . . . have been rejected on First Amendment grounds.”

For other cases in this vein, check out Ari Cohn explaining why a law making social media platforms liable for what posts their algorithms promote is doomed to fail.

Which brings us back to Garcia and the argument FIRE made in our brief — and will inevitably have to make again.

If courts force AI developers to answer in tort every time a user has a tragic or delusional reaction to a chatbot, the incentive structure becomes obvious. They would have to “sanitize their outputs to only the most safe, anodyne, and bland ideas fit for the most sensitive members of society.” In other words, unless you want BarneyBot to be the only AI you’re allowed to use, think twice about demanding that developers anticipate the actions of every conceivable kind of user.

But it’s even worse than that. Movies and music are to a large extent statically consumed. AI helps people create and speak. It’s not only a question of what content AI can deliver to you, it’s a matter of what you will be able to say using AI. Total safety tends to come at a steep — and unacceptable — price.

When the Trump administration demanded changes to Anthropic’s AI system and backed it up with a threat to seize the system or blacklist the company, the message was clear: comply or be crushed. But cut through the rhetoric and the real question is whether Washington can bankrupt a company for saying no to the Pentagon.

Subscribe now

Though the media is busy framing this as a national security showdown, it actually poses a constitutional concern. It is a test of whether the federal government can weaponize its contracting power to force a private company to bend the knee.

AI systems are powerful expressive systems. They generate language, shape ideas, consume and interpret knowledge, and embody the values embedded in their design. The developer has the right, protected by the First Amendment, to make decisions about what capabilities to include or exclude. Anthropic isn’t willing to remove safeguards from its models for use in autonomous weapons targeting or domestic surveillance. Those limits reflect a deliberate expressive choice about what tools the company is willing to build, including what it is willing to provide to the government, and what its existing AI system is capable of achieving.

According to reports, on Feb. 24, War Secretary Pete Hegseth demanded Anthropic CEO Dario Amodei allow unrestricted use of its models “for all legal purposes” within three days or face severe consequences. Those consequences reportedly included blacklisting the company or invoking a Korean War-era law, the Defense Production Act (DPA), to take control of the company’s technology.

Anthropic refused.

In a statement, the company reiterated its commitment to responsible deployment and strict usage policies. Amodei later said the company had simply exercised its “classic First Amendment rights to speak up and disagree with the government.”

The administration’s response was swift. President Trump directed federal agencies to cease using Anthropic’s technology. Secretary Hegseth announced that “no contractor, supplier, or partner that does business with the United States military may conduct any commercial activity with Anthropic.”

Yesterday, the Department of War officially informed Anthropic’s leadership that the company and its products are deemed a supply chain risk, effective immediately.

The government’s actions, which are designed to harm Anthropic’s business, raise serious constitutional concerns, including threats of compelled speech and retaliation against a company for taking positions disfavored by government officials.

First, compelled speech. Anthropic’s decision to build specific guardrails stems from a principled disagreement about how its tools should be designed and used. The company has drawn a line against mass domestic surveillance, warning that AI can assemble commercially available data about Americans’ movements, browsing, and associations into detailed profiles at massive scale, posing serious risks to civil liberties. It has also declined, for now, to power fully autonomous weapons, arguing that today’s systems are not reliable enough to make life-and-death targeting decisions without human oversight.

Forcing Anthropic to remove those limits would compel the company to design and generate capabilities it affirmatively rejects, and has not contracted with the government to provide. And, thankfully, the First Amendment prohibits the government from forcing private speakers like Anthropic to create speech they oppose. Whether it’s a printed pamphlet or coding to enable autonomous targeting, the principle is the same.

When Hegseth threatened to invoke the Defense Production Act to take control of an AI system, he threatened the company with a clear message: The Pentagon is willing to use extraordinary powers to get its way. Enacted during the Korean War, the DPA was designed to mobilize industrial production for national defense, allowing the government to prioritize contracts and direct the manufacture of critical goods. In recent years, its use has expanded beyond traditional wartime manufacturing into domestic production and infrastructure, most notably during the pandemic, when it was invoked to accelerate the production of medical supplies.

Had it actually transpired here, applying the DPA to AI would risk giving the state control of knowledge production. Had Hegseth gotten his way, the government would have overridden the design, training, and limits that reflected Anthropic’s expressive choices about its model.

Second, retaliation. Labeling a domestic company as a security risk is an unprecedented move, and comes immediately on the heels of Anthropic’s refusal to alter its idea of what a responsible AI model should look like. When the government deploys extraordinary coercive power — particularly when justified on emergency or national security grounds — to punish a company for refusing to bow to the government’s demands, the line between legitimate procurement decisions and unconstitutional retaliation grows dangerously thin.

If the government can weaponize contracts and national security laws to coerce companies to reshape their AI systems, developers across the industry will rationally feel pressure to conform their research and design choices to official priorities. Systems guided by independent ethical guardrails risk becoming instruments of state policy via government takeover.

Of course, there are real concerns about national security and foreign adversaries outpacing the United States in AI and military development. But constitutional limits don’t evaporate in these moments. In fact, they matter more when the stakes are high and the pressure is on to centralize government power.

If the government wants AI systems without Anthropic’s restrictions, it can develop its own or contract with companies willing to provide them. (Indeed, OpenAI reportedly stepped in quickly as a replacement.) What the government cannot do is coerce a private company into abandoning its own design principles, or punish it for refusing. That violates the First Amendment.

This essay was originally published in The Dispatch on Feb. 26, 2026.

“The FTC is not the speech police,” FTC Chairman Andrew Ferguson wrote to Apple CEO Tim Cook on Feb. 12. Ferguson might want to ask himself why he felt compelled to pre-empt such an accusation.

Subscribe now

The letter cited reports claiming Apple News “has systematically promoted news articles from left-wing news outlets and suppressed news articles from more conservative publications.” He warned Cook that this may put Apple in violation of its own terms of service as well as the Federal Trade Commission Act — a 1914 federal law establishing the FTC and preventing unfair methods of competition and deceptive acts in commerce — for “material misrepresentations and omissions” toward its consumers.

“I encourage you to conduct a comprehensive review of Apple’s terms of service and ensure that Apple News’ curation of articles is consistent with those terms, and, if it is not, to take corrective action swiftly,” Ferguson concluded.

This kind of baseless, flagrant intimidation tactic would be frightening if it weren’t so obviously ridiculous.

For starters, the claim about left-wing bias comes from recent research released by the Media Research Center. According to MRC, “only one of the 560 articles examined in November came from a right-leaning source.” But the reliability of MRC data is far from a given. To begin with, it relies on subjective pronouncements of what outlets are “left-wing,” which are at best arbitrary and at worst themselves biased. Critics have also noted that some of the news outlets the MRC claimed were suppressed, such as Breitbart and The Gateway Pundit, don’t even post on the platform. Yes, really. And on top of it all Apple News is an aggregator with in-depth customization tools. Users can tailor the platform to their own preferences, and they can even eliminate all of Apple’s editorial content if they so choose. Writing for Forbes this month, media and culture journalist Andy Meek concluded that “the app largely reflects the choices that its users make themselves.”

But even if the allegations were true, they would still be irrelevant.

The fact is that Apple has the same First Amendment right to decide which content to share or feature as a newspaper editor does. The government isn’t free to replace Apple’s editorial judgment with its own just because it disagrees with Apple’s decisions. Imagine Ferguson asserting authority over the order and visibility of newspapers and magazines at newsstands, or the organization of bookstore shelves. Everyone would view that as unacceptable government overreach, and there is no reason to view it otherwise simply because the target is digital.

Or imagine the FTC investigating whether Fox News was really “fair and balanced,” as its slogan used to claim, threatening punishment for unfair or deceptive trade practices if the government decided its coverage was slanted. Actually, you don’t have to imagine it — the FTC received a petition asking it to do exactly that, which the agency properly rejected as obviously foreclosed by the First Amendment.

Apple’s First Amendment rights aren’t limited by the fact it’s a technology company, despite the FTC’s attempt to frame it as an issue of the platform’s terms of service. This kind of inquest is about as transparent and lazy as censorial pressure can get. But the consequences of letting it slide can be fatal to free speech and a liberal democracy. If the government can supplant a website or app’s interpretation of their subjective terms of service with its own, the FTC could easily censor any ideas it doesn’t like by threatening to punish an alleged “failure to live up to” them.

Democratic administrations could use “terms of service” inquiries to pressure private platforms and organizations into eliminating “hate speech.” Conservative administrations could do the same to force removal of posts calling them “fascists” or reporting on ICE activity in public.

It’s precisely this kind of “We must violate the First Amendment in order to save it” tactic that led Supreme Court Justice Elena Kagan to conclude in Moody v. NetChoice: “On the spectrum of dangers to free expression, there are few greater than allowing the government to change the speech of private actors in order to achieve its own conception of speech nirvana.”

This is just the latest example of the coercive and censorial abuses of power the FTC has taken during President Trump’s second term. In May 2025, it opened an investigation into the liberal watchdog organization Media Matters for America, claiming illegal collusion with other advocacy groups to cut off X’s advertising revenue.

Media Matters sued the FTC in June, arguing that the agency’s action was retribution for a 2023 report showing that ads on X for brands like Apple, IBM, and Xfinity were shown alongside antisemitic content. In August, a federal judge blocked the investigation, calling it “a straightforward First Amendment violation.” As the judge ruled, Media Matters did nothing more than engage in “quintessential First Amendment activity when it published an online article criticizing Mr. Musk and X.”

The FTC is waging a similarly censorial campaign against NewsGuard, a private news and information website rating company, over its alleged practice of maliciously rating conservative media companies like Newsmax lower than liberal ones like the New York Times. Yet NewsGuard rates both National Review and the Wall Street Journal, both considered right of center, higher than the New York Times. That, of course, didn’t stop the FTC. Claiming antitrust violations, the FTC has demanded that NewsGuard turn over virtually every document it has created or received since its founding in 2018 — including memos, emails, texts, reporters’ notes, subscriber lists, analyses, financial reports, and more.

And you thought an IRS audit was bad.

But that’s not all. The FTC also forced other private companies to blacklist NewGuard. When the large media corporation Omnicom announced a merger with Interpublic, the FTC immediately launched an investigation. Their claim was that the merged entity — which would become the largest advertising entity in the world — could violate competition law by steering “advertising away from media publishers based on the publishers’ political or ideological viewpoints.” In other words, by exercising their First Amendment rights. As a condition for approving the merger, the FTC barred the newly formed company from using NewsGuard’s ratings and journalists — which it has been doing as early as March 2021.

Once again, this is a clear violation of the First Amendment. That’s why, with the help of our organization, the Foundation for Individual Rights and Expression (FIRE), NewsGuard is suing the FTC for this unconstitutional speech-based retaliation and jawboning.

While Ferguson’s letter to Apple correctly notes that he and the FTC “do not have authority to require Apple or any other firm to take affirmative positions on any political issue, nor to curate news offerings consistent with one ideology or another,” that’s exactly what he’s doing. The FTC has no more business policing the editorial decisions of Apple News under the guise of “consumer protection” than it does the “fair and balanced-ness” of Fox News’ coverage.

Ferguson’s letter is egregious and absurd, and one can only hope that Apple will push back on this pressure — especially since Tim Cook’s attempts to flatter and gift his way out of the crosshairs didn’t seem to work for very long. The response to these strong-arm tactics is principled opposition and a willingness to stand on its rights. We’re lucky enough to have these legal tools to fight back with, but they’re nothing without the spine and resolve needed to wield them.

When the history of this peculiar period is written, the story will be told not by those who capitulated, but by those who chose to defend time-tested constitutional principles.

Robert Corn-Revere is chief counsel of the Foundation for Individual Rights and Expression as well as former chief counsel to Federal Communications Commission Chairman James H. Quello.

The president doesn’t like it when late-night TV talk show hosts make fun of him.

That undeniable fact is the common denominator for a series of actions that have kept the Federal Communications Commission in the news this past year and why Brendan Carr, its chairman, has been talking lately about the FCC’s “equal time” rule.

Subscribe now

The controversy over the so-called equal time rule surfaced for most of the public recently when Stephen Colbert said CBS refused to allow him to air his interview with Texas Democratic Senate candidate James Talarico after the FCC signaled it would enforce the rule against TV talk shows. (For the record, CBS disputed his account and said the network had only “provided legal guidance” for how the show could comply with the rule. Nice show you got here . . .)

Colbert opted to do the interview and release it on The Late Show’s YouTube channel (where FCC rules do not apply) but not air it on the network or its affiliated stations. And in a perfect demonstration of the Streisand effect, the interview almost immediately racked up over 7.5 million views, triple the nightly broadcast audience for The Late Show.

There are many reasons why the FCC’s equal time rule is obsolete (and likely unconstitutional in today’s technological environment) and why Carr’s reliance on it to squelch television talk show interviews is wildly inappropriate. But first, let’s connect the dots, shall we?

The president has never been shy about expressing his displeasure with talk show hosts who roast him, calling them out by name in a seemingly endless stream of Truth Social posts, crowing at the announcement that Colbert’s show would not be renewed (mysteriously in the wake of FCC approval of the Skydance Media-Paramount Global merger), and predicting that Jimmy Kimmel would be the next to go.

And Carr, President Trump’s hand-picked chairman, has turned himself inside out looking for ways to do his master’s bidding. He quite literally has inverted himself: He used to say the FCC cannot act as the nation’s speech police, while now he can’t wait to pin on a badge — in this case, a Donald Trump lapel pin — and patrol “his” beat.

He impersonated a Mafia don in threatening ABC over a Kimmel monologue (“we can do this the easy way or the hard way”). In the process, Carr has made reference to a series of different FCC policies in various efforts to muzzle TV comics, none of which can be taken seriously by anyone who understands the Communications Act or the First Amendment.

Carr frequently cites the “public interest” requirements imposed on broadcast licensees, but that general obligation to serve the public (that is, to provide broadcasting service) has never been understood to give the FCC power to restrict particular programs, and it would be unconstitutional if it did. He neglects to note that both the FCC and the courts historically have interpreted the public interest standard to require broadcasters to exercise independent editorial judgment, and not to follow the whims of some government functionary.

Carr has even tried to summon the “news distortion rule” as justification to threaten networks over late-night comedy monologues, as if it ever had any conceivable application in that context. But that long-moribund rule for news programs, which the FCC unearthed to go after 60 Minutes for the way it edited its Kamala Harris interview, does not give the FCC the power to sit as editor in chief. And any attempt to enforce it in court would instantly be declared a First Amendment violation.

But who has to worry about courts if you can bully broadcasters into submission?

All Carr’s casting about for an excuse to lean on programming the president dislikes brings us to the latest gambit, the “equal time” rule.

More accurately, as set forth in the Communications Act, it is the “equal opportunity” rule, and it is triggered when a broadcaster allows a “legally qualified” candidate to “use” the station — that is, to appear on air. Once that happens, the station must offer comparable opportunities to opposing candidates but only if they request it. But Congress added exemptions to the rule in 1959, including for news interviews, and for decades the FCC applied those exemptions broadly to encourage the widespread discussion of political affairs. This included exemptions for candidate interviews on the radio, including for The Howard Stern Show, and since at least 2006, the FCC held that late-night shows qualified for that exemption.

But the FCC’s Jan. 21, 2026 guidance for the first time narrows the exemptions, telling broadcasters they should no longer assume that their late-night shows qualify for it. From Carr’s perspective, it is just another regulatory tool that can be cynically employed to squeeze TV hosts with the temerity to make fun of his boss.

This departure from its traditionally broad reading of the exemptions is far from neutral. The Commission framed its January 2026 guidance as applying to “broadcast television stations,” and as Carr put it last month (channeling his master’s voice), if you are “fake news,” you won’t qualify for the news interview exemption. But at the same time Carr suggested that talk radio won’t be affected by this reinterpretation, signaling that this is about selective pressure on stations the Trump administration disfavors. Translation: Carr’s FCC won’t be going after Glenn Beck or Dan Bongino for their talk radio shows anytime soon.

This is an object lesson in why it is a bad idea to vest the federal government with discretionary power over an important medium of communication. Back in the day (the 1980s), two principled Republican FCC chairmen, Mark Fowler and Dennis Patrick, took action during the Reagan administration to eliminate the Fairness Doctrine, which required broadcasters to present opposing viewpoints on controversial issues of public importance. The Fairness Doctrine was based on the same theories used to justify the equal-time rule (as well as the news-distortion rule), but Fowler and Patrick maintained that past justifications for broadcast regulation were no longer valid, and that broadcasters should receive the same First Amendment guarantees that the print media have enjoyed since our country’s inception.

Interesting side note: Last November, a bipartisan group that included seven former FCC commissioners (five Republicans and two Democrats), four of whom had served as chair (including Fowler and Patrick), filed a petition with the FCC asking that the Commission cease its partisan manipulation of broadcast rules and to eliminate the news-distortion rule. Carr’s reaction? Less than a day after the filing, he sneered at the petition in a social media post: “How about no.” Reasoned decisionmaking, indeed.

But the former FCC officials are correct — it is long past time to eliminate broadcast content controls that were adopted under different technological conditions when broadcasting was the only electronic mass medium in town. And, as the fiasco over the equal time rule and Colbert’s deft circumvention demonstrate, political speech isn’t confined to licensed channels. As was once said of cyberspace, “the internet treats censorship like system damage and routes around it.”

Just so. And here, the damaged system is the FCC.

Broadcasters can distribute content outside of the FCC’s jurisdiction, most notably, to the internet. In a world where political speech is abundant and distributed across countless channels and platforms, the spectrum scarcity rationale that propped up broadcasting regulation has taken on far too much constitutional water to stay afloat.

As the former FCC commissioners’ petition to eliminate the news-distortion rule demonstrates, this is not a partisan issue. And the current crop of Republicans aren’t the only ones who have tried to aim the FCC’s licensing power at disfavored broadcasters. In 2018, a group of Senate Democrats urged the FCC to investigate Sinclair Broadcasting — which The New Yorker once called a “conservative media empire” — when Sinclair sought to merge with Tribune Media Company. The letter said that Sinclair violated its “public interest obligation” by “staging, slanting, or falsifying information” and suggested that the FCC “could disqualify Sinclair from holding its existing licenses” or obtaining new ones.

The tendency to leverage the FCC’s licensing power against hostile broadcast news coverage is nothing new. During the 1960s, the Kennedy administration quietly encouraged the FCC to scrutinize conservative radio broadcasters under the “public interest” standard and the fairness doctrine in an effort to blunt its critics. And Richard Nixon famously sought to use the FCC to target broadcast stations owned by the Washington Post because of the newspaper’s Watergate coverage.

The moral of this story is pretty simple: The government cannot be trusted with this kind of power, regardless of which party is in office.

Debates over free expression often center on government power and the First Amendment. But in fandom communities and other niche online subcultures, the boundaries of speech are shaped by moderators, platform policies, and evolving group norms. Within these intensely participatory spaces, decisions about what is acceptable can determine which voices are amplified and which are pushed aside. In these environments, cultural gatekeeping and platform rules often define who gets heard.

Subscribe now

When Arthur Conan Doyle killed off Sherlock Holmes in 1893, fans lost their minds. They wrote angry letters to Doyle and his publisher, they wore black armbands in the streets as if a real person had died, they even began writing and publishing their own unauthorized stories about the beloved detective. The practice of fanfiction is as old as storytelling itself. The Aeneid builds on Homer’s Iliad, Shakespeare’s Romeo and Juliet reimagines a poem by Arthur Brooke, and Dante’s Inferno is sometimes described as “self-insert” fanfiction of the Bible. But like any form of artistic expression, fanfic has long faced creative restraints.

In 1981, for example, Lucasfilm sent letters to Star Wars fanzine publishers saying they were free to continue — so long as they didn’t include pornographic stories. Nor has moving from print to the internet resulted in a landscape free from constraints. Often, fanfic platforms remove content without warning based on input by platform owners, advertisers, or the public. FanFiction.Net purges adult content despite having an “M” rating for mature stories, and has banned entire genres including self-inserts, scripts, songfics, and audiofics (“fic” meaning a work of fanfiction). The platform Wattpad has removed LGBT stories from the Warrior Cats series after parents complained and targeted LGBT content more broadly.

Recognizing the risks to free expression, fans founded Archive of Our Own in 2008 with a clear mission: “Maximum inclusiveness of fanwork content.” Unlike other fanfic platforms, AO3 maintains a strong opposition to creative restrictions. The site imposes only two content requirements: all work must be fan-made and users cannot claim other people’s work as their own. Based in the United States, AO3 now serves more than 10 million users, supports dozens of languages, and hosts over 16.7 million works across more than 76,000 fandoms. Time magazine named AO3 one of the 50 Best Websites of 2013. And in 2025, Forbes listed it as one of the world’s best fanfic sites, alongside FanFiction.Net.

To help users avoid content they don’t want to see, AO3 offers a robust tagging system with ratings and content warnings, allowing readers to filter or mute specific themes while authors can choose to post works without rating them at all.

But sexual content is just one common target. Within fandom communities, calls for censorship are growing louder, with other targets including depictions of racism or other forms of discrimination, abuse, violence, or underage characters dealing with “adult” topics.

Once-common fandom maxims like “don’t like, don’t read” or “ship and let ship” (let bygones be bygones, but applied to character relationships, or “ships”) have given way to claims that depicting harmful behavior in fiction necessarily encourages it in real life. This mirrors a broader cultural trend FIRE has written about — the collapsing distinction between words and violence.

In some cases, fans have even gone so far as to reinvent something akin to the Hays Code, a Hollywood self-censorship rulebook that was in use from the 1930s to 1960s, instructing early moviemakers on how to avoid offending America’s moral watchdogs. Fans have argued, as Hays did in its time, that depictions of morally questionable behaviors are only okay if they are punished within the story. An adulterous character must see the error of their ways. A villain must face consequences. The abuser cannot be portrayed sympathetically, even for a chapter.

The underlying concerns driving these restrictions aren’t entirely baseless — research on media effects shows that repeated exposure to certain content can normalize attitudes, particularly among younger audiences. And platforms do face real legal and ethical questions about hosting mature material, even if fictional. But AO3 argues that categorical content bans only create more problems than they solve. In its FAQ, it explains, “Biased enforcement of content rules has been shown to occur even when the purpose of the rule is to push back against discrimination. For example, rules intended to reduce racial hate speech on social media often end up being disproportionately enforced against racial minorities speaking out against racism.”

They’re not wrong. In fact, this has played out repeatedly. In 2018, for example, Tumblr’s adult content ban disproportionately flagged art by black creators while missing actual pornography.

Fans also point out that writing about darker themes like child abuse, racial discrimination, or sexual assault can be cathartic for survivors. Psychologists have found that journaling and creative writing can help people process trauma, and such practices are often used in therapy for PTSD and related conditions. When asked why it doesn’t remove extremely offensive content, AO3 offers a blunt defense of creative freedom:

Our mission is to host transformative fanworks without making judgments based on morality or personal preferences. If it’s a fictional fanwork that is legal to post in the United States, then it is welcome on AO3. This approach is intended to reduce the risk that content will be removed as a result of cultural or personal bias against marginalized communities.

We recognize that there are works on AO3 that contain or depict bigotry and objectionable content. However, we are dedicated to safeguarding all fanworks, without consideration of any work’s individual merits or how we personally feel about it. We will not remove works from AO3 simply because someone believes they are offensive or objectionable.

In fandom communities, the forces shaping speech are platform policies and ever-evolving community norms. Unlike the constitutional clashes that define disputes over government power, these conflicts play out in message boards and comment threads. The stakes may appear smaller than heavy-handed government regulation, but for the people involved, they shape who gets to participate and what ideas are allowed to take root.

AO3 has made its website a bulwark in an online landscape increasingly shaped by censorship and moral panic, distinguishing itself as a lighthouse in the storm. As a private platform, it retains the right to set and enforce its own rules, just as users remain free to express themselves. That tension between platform discretion and user expression may not be a constitutional crisis, but it is a reflection of how cultural and platform values shape today’s digital spaces.

Last Monday, FBI Director Kash Patel announced an investigation into Signal group chats that Minnesotans are using to track ICE activity. Independent journalist Cam Higby spurred the move with an X thread that appears to show users of the encrypted messaging app reporting ICE sightings and sharing license plate numbers of agency vehicles. What the thread doesn’t show is evidence of a crime.

Subscribe now

Patel claimed sharing such information is illegal if it “leads to a break in the federal statute or a violation of some law,” adding, “you cannot create a scenario that illegally entraps or puts law enforcement in harm’s way.” Border czar Tom Homan sounded even more certain. Asked about the chats later that week, he said, “I’m not going to show our hand. But they’ll be held accountable. Justice is coming.”

But speech does not lose constitutional protection simply because it might lead others to break the law. That was true when progressive commentators warned about “stochastic terrorism” — the idea that conservative rhetoric on hot-button issues incites violence against minority groups — and it’s true now. There isn’t even evidence in the leaked Signal chats that anyone did use the information to commit a crime.

Consider the relevant First Amendment exceptions. True threats are serious expressions of intent to physically harm a specific person or group. Incitement is speech intended and likely to produce imminent lawless action. Conspiracy consists of an agreement to commit a specific crime and an overt act toward carrying it out. Aiding and abetting involves intentionally and substantially assisting a specific criminal act. None of these categories covers the mere sharing of information that others can use — and have been using — for lawful purposes, such as protesting, observing, or documenting public law enforcement activity. Higby’s X thread shows nothing more.

As the Supreme Court put it, “The freedom of individuals verbally to oppose or challenge police action without thereby risking arrest is one of the principal characteristics by which we distinguish a free nation from a police state.”

Of course, anyone who assaults a federal agent or physically interferes with an enforcement operation can and should be prosecuted. But, absent evidence of conspiracy or aiding and abetting, as those terms are actually defined under the law, that crime does not retroactively strip speech of First Amendment protection. Google Maps isn’t culpable if someone uses it to vandalize an ICE facility or an abortion clinic.

It’s possible to imagine circumstances in which anti-ICE activists’ speech would lose constitutional protection. For example, if two people share an ICE agent’s whereabouts and agree to meet there to assault the agent, then start taking action toward committing that crime, they would be guilty of conspiracy. Outside such narrow circumstances, however, the First Amendment protects sharing information about law enforcement, much as millions of drivers do every day when they report police locations on apps like Waze.

These First Amendment exceptions are narrow and precise by design. They capture a sliver of speech that is inseparable from criminal conduct, without giving the government sweeping power to suppress dissent.

The FBI’s investigation fits a broader pattern. The Trump administration has repeatedly threatened to go after Americans for protesting, monitoring, or speaking about immigration enforcement. Officials frame these threats as crackdowns on “doxxing,” “impeding,” or “obstructing” federal agents. What they’re actually doing is taking words that sound like they describe crimes and quietly stretching their meanings until they cover a wide range of protected activity, hoping that the scary labels will blunt any pushback or skepticism.

This tactic is an example of what my colleague Angel Eduardo calls “linguistic parasitism” — the “stealth-redefinition or expansion of a word, phrase, or concept’s meaning while seizing upon its common meaning to elicit the desired response.” But this administration isn’t eliciting the desired response from civil libertarians. Every time an official says “doxxing” or “impeding,” I hear the voice of Inigo Montoya: “You keep using that word. I do not think it means what you think it means.” When you drill down, you realize these accusations often refer to activities like filming ICE agents and posting photos and videos of them online.

A month after President Trump’s inauguration, Homan asked the Justice Department to investigate Rep. Alexandria Ocasio-Cortez for “impeding” law enforcement by releasing a webinar and flyer explaining people’s constitutional rights during ICE encounters. Last July, after CNN reported on ICEBlock — an app that lets users report ICE sightings due to concerns over the agency’s “alleged civil rights abuses and failures to adhere to constitutional principles” — Homan again urged DOJ to investigate whether CNN was illegally impeding law enforcement by reporting on the app. ICEBlock itself later disappeared from the App Store, and Attorney General Pam Bondi acknowledged that the DOJ has “demanded” the tech company remove it — a textbook example of jawboning.

In August, ICE tagged the Department of Justice in a repost of Libs of TikTok’s post accusing Connecticut Rep. Corey Paris for “doxxing ICE’s live location” and demanding prosecution. What had Paris done? He announced on Instagram that he received reports of ICE activity in his district and urged residents to “remain vigilant” and “seek out trusted legal and community resources if needed.” Paris ultimately was not charged with a crime for noting that law enforcement activity was taking place somewhere in a 2.5-square-mile area.

Given this pattern of threats and rhetoric, it’s no surprise that incidents keep emerging in which federal agents confront and threaten protesters and observers for exercising their First Amendment rights. In one recent video, a masked ICE agent told a woman recording him that he was photographing her car because “we have a nice little database and now you’re considered a domestic terrorist.”

Maybe the FBI’s Signal investigation will quietly fade away. But the chilling effect will remain. It was bad enough when, during Joe Biden’s presidency, the FBI pressured social media companies to censor protected speech deemed dangerously misleading. Now the bureau is treating protected speech on an encrypted messaging app as grounds for criminal investigation.

This isn’t just about opposition to ICE. It’s about the right of every American to criticize, discuss, protest, observe, and document what the government is doing, regardless of who is in power or what the cause is.

The government can punish violence. It can punish actual obstruction. What it cannot do is erase the line between criminal conduct and free speech. Once that line disappears, no one’s rights are safe.