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.

Subscribe now

Government efforts to mandate their preferred ideology risk replacing one orthodoxy with another, and laws aimed at preventing “algorithmic discrimination” are already punishing uncomfortable facts. As FIRE has warned, this is the return of speech codes — not just on campus, but in code itself. If we get this wrong, we risk enshrining conformity at the very source of knowledge and silencing the dissent that truth depends on.

In the first episode of Free Speech Future, recorded Sept. 18, 2025, a panel of builders and defenders confronts the risk that emerging technologies will be optimized for harmony at the expense of truth.

• Andrew Mayne (X), GP Zero Shot Fund and host of the OpenAI Podcast

• Greg Lukianoff (X), president and CEO of FIRE

• Brendan McCord (X), founder of Cosmos Institute

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

  Expression is a reader-supported publication. To receive new posts and support our work, consider becoming a free or paid subscriber.

Grok, the AI system integrated into X, has reportedly been used to turn real pictures of people — including minors — into nude or sexualized imagery.

Subscribe now

People are understandably outraged. This episode shows how a person can use AI tools to violate a sense of human dignity and security with little more than a photo and a prompt. The fact that the tool was used to target real people, especially children, without their knowledge or consent is particularly disturbing to many.

Some have responded by calling for new laws. That instinct is understandable. But many proposals would raise serious First Amendment concerns, and before trying to scratch the “do something” itch with new legislation, it’s important to first ask: does existing law already prohibit this?

In many cases, the answer is yes.

Federal criminal law prohibits knowingly making or sharing child sexual abuse material involving actual children, whether it is created by a camera or with the assistance of AI. Likewise, AI-generated material that meets the high bar for obscenity and is publicly created or distributed, is not protected speech. Users who knowingly prompt an AI system to create such content, or who share it, can already face criminal prosecution. Liability shields don’t protect anyone from federal criminal prosecutions. AI operators that knowingly provide substantial assistance to those creating this unlawful content may face legal exposure as well.

Existing law also provides other avenues to hold people accountable through private lawsuits. Civil claims for harms like intentional infliction of emotional distress, invasion of privacy, defamation, and misappropriation of likeness may also be available to people depicted in the images created by Grok, provided the elements of those torts, and any constitutional protections built into them, are satisfied. These types of claims allow victims to collect monetary damages against users who make, share, or sell such content and, in limited cases, developers.

At the same time, it’s important to be clear about the limits of the law. The law will never be able to fully prevent bad actors from doing bad things. And the Constitution limits how far the government can go in trying. Nudity and sexual content involving adults are generally protected by the First Amendment unless they fall into a narrow category of unprotected speech. Use of AI does not change that constitutional analysis. This means a great deal of offensive or distasteful expression remains protected speech, even when it disturbs or makes us uncomfortable.

This matters. If every technological failure becomes an excuse to expand government authority over speech, the predictable outcome is overreach that chills expression and silences voices.

Public pressure, reputational risk, and the possibility of lawsuits are powerful incentives to motivate xAI, the parent company of both Grok and X, to improve safeguards, redesign systems, and limit misuse. That is the preferred path. Editorial and design decisions made by private companies are far less dangerous than granting the government broad power to regulate speech and assume control over platforms protected by the First Amendment.

Using Grok’s failures as a justification for sweeping new AI speech regulations would be a mistake. Existing laws already target real harms and real actors. Broad new rules risk overreach, chilling lawful expression and empowering the state in ways that are difficult to unwind.

The right response here starts with enforcing the law we already have, and to resist the temptation to trade constitutional principles for the illusion of control.

As lawmakers kick off the 2026 legislative session, a new and consequential phase in the conversation about free speech and artificial intelligence is already taking shape in statehouses across the country. Yet another crop of AI bills is set to dictate how people use machines to speak and communicate, raising fundamental constitutional questions about freedom of expression in this country.

Subscribe now

The First Amendment applies to artificial intelligence in much the same way it applies to earlier expressive technologies. Like the printing press, the camera, the internet, and social media, AI is a tool people use to communicate ideas, access information, and generate knowledge. Regardless of the medium involved, our Constitution protects these forms of expression.

As lawmakers revisit AI policy in 2026, it bears repeating that existing law already deals with many of the harms they seek to address — fraud, forgery, defamation, discrimination, and election interference — whether or not AI is used. Fraud is still fraud, whether you use a pen or a keyboard, because liability properly attaches to the person who commits unlawful acts rather than the instrument they used to do it.

Many of the AI bills introduced or expected this year rely on regulatory approaches that raise serious First Amendment concerns. Some would require developers or users to attach disclaimers, labels, or other statements to lawful AI-generated expression, forcing them to serve as government mouthpieces for views they may not hold. FIRE has long opposed compelled speech in school, on campus, and online, and the same concerns apply when it comes to AI systems.

Election-related deepfake legislation remains a central focus in 2026. Over the past year, multiple states have introduced bills aimed at controlling AI-generated political content. But these laws often restrict core political speech, and courts have applied well-settled First Amendment jurisprudence to find them unconstitutional. For example, in Kohls v. Bonta, a federal district court struck down California’s election-related deepfake statute, holding its restrictions on AI-generated political content and accompanying disclosure requirements violated the First Amendment. The court emphasized that constitutional protections for political speech, including satire, parody, and criticism of public officials, apply even when new technologies are used to create that expression.

Another growing category of legislation seeks to restrict “chatbots,” or conversational AI, using frameworks borrowed from social media laws. These include blanket warning requirements telling users they are interacting with AI, sweeping in many ordinary, low-risk interactions where no warning is needed. Some proposals would categorically prohibit chatbots from being trained to provide “emotional support” to users, effectively imposing a direct and amorphous regulation on the tone and content of AI-generated responses. Other proposals require age or identity verification, either explicitly or as a practical matter, before a user may access the chatbot.

These kinds of constraints place the government between the people and information they have a constitutionally protected right to access. They censor lawful expression and burden the right to speak and listen anonymously.

For that reason, courts have repeatedly blocked similar restrictions when applied to social media users and platforms. The result is likely to be similar for AI.

Broad, overarching AI regulatory bills have also returned this year, with at least one state introducing such a proposal so far this cycle. These bills, which were introduced in several states in 2025, go well beyond narrow use cases, seeking to impose sprawling regulatory frameworks on AI developers, deployers, and users through expansive government oversight and sweeping liability for third-party uses of AI tools. When applied to expressive AI systems, these approaches raise serious First Amendment concerns, particularly when they involve compelled disclosures and interfere with editorial judgment in AI design.

Addressing real harms, including fraud, discrimination, and election interference can be legitimate legislative goals. But through FIRE’s decades of experience defending free expression, we’ve observed how expansive, vague, and preemptive regulation of expressive tools often chills lawful speech without effectively targeting misconduct. That risk is especially acute when laws incentivize AI developers to suppress lawful outputs, restrict model capabilities, or deny access to information to avoid regulatory exposure.

Rather than targeting political speech, imposing age gates on expressive tools, or mandating government-scripted disclosures, government officials should begin with the legal tools already available to them. Existing laws provide remedies for unlawful conduct and allow enforcement against bad actors without burdening protected expression or innovation. Where gaps truly exist, any legislative response should be narrow, precise, and focused on actionable conduct.

American universities insist they are committed to free expression. Many now point to “dialogue-across-differences” programs and AI-mediated conversations as proof. Students are trained to listen respectfully, disagree productively, and engage across differences.

I see these initiatives up close: I sit in faculty meetings where they are rolled out, watch students navigate them in real time, and hear privately what they are reluctant to say publicly. And the climate for speech on campus continues to deteriorate.

Subscribe now

According to the Foundation for Individual Rights and Expression’s 2026 College Free Speech Rankings, most American colleges now earn failing marks for speech climate. Of the 257 institutions surveyed, 166 received an F, and only 11 earned a C or higher. The national average score — 58.63 out of 100 — would itself be a failing grade in most classrooms. Student attitudes have hardened as well: Majorities now oppose allowing controversial speakers from across the political spectrum.

Whatever these initiatives are meant to accomplish, the broader data suggest they are not reversing the decline in speech climate.

The problem is not that students lack conversational skills. In my classrooms, students are thoughtful, capable, and often eager to engage. What they lack is confidence the institution will protect them when their viewpoints fall outside prevailing norms. Universities increasingly treat speech as something to be managed rather than protected, and students understand that distinction intuitively.

A recent essay by Hollis Robbins, professor of English at the University of Utah and former dean of humanities there, in the Chronicle of Higher Education helps clarify what is going wrong. In “The Noxiousness of Civic-Discourse Platforms,” Robbins examines the design of popular civic-discourse platforms that require students to disclose their positions on controversial issues, rate their confidence in those views, and participate in moderated dialogues that are peer-evaluated and archived.

The stated goal is openness. The effect is closer to surveillance.

Platforms like Sway — already in use on at least 77 campuses and backed in part by U.S. intelligence community funding — and Dialogues typically work by prompting students to enter their views on contested topics such as abortion, immigration, and policing, then matching them with peers who hold opposing positions for structured online exchanges. What was once a casual conversation becomes part of a permanent record. Beliefs are transformed into data — recorded, scored, and retained. Discretion is limited, and opting out often comes with its own costs. Over time, institutions accumulate detailed records of what students say, how confidently they say it, and how others judge them for it.

That reality shapes behavior long before anyone logs on.

For much of modern higher education, liberal learning rested on a simple norm: Students could explore ideas without being defined by them. They could be uncertain, incomplete, or undecided. That space for intellectual risk was not incidental. It was essential.

Many contemporary discourse initiatives narrow that space. When students are asked to quantify their agreement, are sorted into opposing camps, and evaluated on their expression of belief, they learn quickly that speech is consequential. In such environments, caution is not cowardice. It is rational behavior — and it is one I see play out semester after semester.

This helps explain a pattern FIRE has documented repeatedly: Students report growing reluctance to speak openly, even as universities expand programming supposedly designed to encourage it.

Not all discourse initiatives make these mistakes. Programs that operate in person rather than through digital platforms, and that follow the Chatham House Rule — where participants can share ideas but not attribute them — preserve the exploratory space that liberal learning requires. FIRE’s own Let’s Talk program takes this approach: facilitating face-to-face conversations where what is said in the room stays in the room, and where no algorithm records or scores participants’ beliefs.

These distinctions matter. A conversation governed by discretion and conducted in person is fundamentally different from one where beliefs become data points in a permanent file. The former creates conditions for genuine exploration; the latter creates incentives for performance.

But many platforms gaining traction on campuses abandon both principles. They move dialogue onto screens, mediate it through algorithms, and treat transparency as an unqualified good — missing that discretion and physical presence are often what make honest inquiry possible.

The deeper failure is institutional.

As FIRE has warned, free speech is increasingly treated as a conditional privilege rather than a foundational principle — defended when convenient, constrained when controversial. Universities are often eager to endorse free expression in theory, but hesitant to defend it when speech provokes outrage, reputational risk, or political pressure.

Many dialogue platforms reflect this same logic. They create controlled spaces for managed disagreement while leaving intact — or even reinforcing — the broader culture of fear surrounding unscripted speech. They teach students how to talk without assuring them that they will be protected when they do.

Worse, the turn toward virtual dialogue sidesteps the real challenge universities face: building pluralism where people actually live and learn together.

Free speech is not learned in apps or simulations. It is learned in dorm rooms, dining halls, classrooms, and student organizations — places where people with genuinely different backgrounds, beliefs, and temperaments must coexist. Anyone who teaches knows this. Online debates cannot substitute for the daily work of navigating disagreement face to face.

What universities increasingly avoid is the harder work of governing pluralistic communities. Residential life, dining halls, and classrooms are where disagreements actually surface — often awkwardly, sometimes painfully. Fostering such spaces requires clear expectations, consistent enforcement of viewpoint-neutral policies, and administrators willing to defend expressive rights when doing so is uncomfortable. No platform can substitute for that kind of leadership.

Universities cannot outsource that responsibility.

The College Free Speech Rankings point to another troubling shift consistent with this analysis. A record one in three students now say that using violence to stop a campus speech can sometimes be justified. Growing numbers say that shouting down speakers or disrupting events is acceptable. This is not merely a breakdown of etiquette. It reflects a deeper confusion about what universities are for and what free speech demands of a community committed to learning.

Platforms that prioritize the expression of belief over the development of judgment do little to reverse this trend. They adapt to the erosion of free-speech norms instead of resisting it.

Universities already know how to cultivate the habits free societies require. Faculty do it every day through serious coursework, evidence-based argument, and disciplined disagreement. That work is slower and harder to quantify. It does not produce dashboards. But it does produce judgment.

Free speech cannot survive as a managed outcome. It cannot be reduced to metrics or mediated by machines. And it will not be restored by tools designed to make disagreement frictionless.

What students need is not another platform. They need institutions willing to say, clearly and publicly, that the expression of unpopular ideas is not a problem to be solved but a condition of learning itself. They need faculty who model disagreement as a scholarly practice. And they need administrators who will defend them when the cost of doing so is real.

The answer to a failing speech climate is not better software. It is better leadership — exercised in real spaces, among real people, facing real disagreement, with real conviction.

With so much of our national conversation taking place online, there’s an almost reflexive tendency to search for online causes — and online solutions — when tragedy strikes in the physical world. The murder of Charlie Kirk was no exception. Almost immediately, many (some in good faith, and others decidedly less so) began to postulate about the role played by online rhetoric and polarization.

Subscribe now

Taking the stage at Utah Valley University to discuss political violence last week, Sens. Mark Kelly and John Curtis shared the view that social media platforms are fueling “radicalization” and violence through their content-recommendation algorithms. And they previewed their proposed solution: a bill that would strip platforms of Section 230 protections whenever their algorithms “amplify content that caused harm.”

This week, the senators unveiled the Algorithm Accountability Act. In a nutshell, the bill would require social media platforms to “exercise reasonable care” to prevent their algorithms from contributing to foreseeable bodily injury or death, whether the user is the victim or the perpetrator. A platform that fails to do so would lose Section 230’s critical protection against being treated as the publisher of user-generated content — and injured parties could sue the platform for violating this “duty of care.”

The debate over algorithmic content recommendation has been going on for years. Lower courts have almost universally held that Section 230 immunizes social media platforms from lawsuits claiming that algorithmic recommendation of harmful content contributed to terrorist attacks, mass shootings, and racist attacks. When faced with the question in 2023, the Supreme Court declined to rule on the scope of Section 230 — opting instead to hold the claims of algorithmic aiding and abetting at issue would not survive either way.

But there’s an important question that usually gets lost in the heated debate over Section 230: Would such lawsuits be viable even if they could be brought?

In a Wall Street Journal op-ed making the case for his bill, Sen. Curtis wrote, “We hold pharmaceutical companies accountable when their products cause injury. There is no reason Big Tech should be treated differently.”

At first blush, this argument has an instinctive appeal. But it ultimately dooms itself because there is a reason to treat social media platforms differently. That reason is the First Amendment, which enshrines a constitutional right to free speech — a protection not shared by prescription drugs.

Perhaps anticipating this point, Sen. Curtis argues that the Algorithm Accountability Act poses no threat to free speech: “Free speech means you can say what you want in the digital town square. Social-media companies host that town square, but algorithms rearrange it.” But free speech doesn’t only protect users’ right to post online free of government censorship; it also protects the editorial decisions of those that host those posts—including algorithmic “rearranging,” to use the senator’s phrase. As the Supreme Court recently affirmed in Moody v. NetChoice:

When the platforms use their Standards and Guidelines to decide which third-party content those feeds will display, or how the display will be ordered and organized, they are making expressive choices. And because that is true, they receive First Amendment protection.

The “rearranging” of speech is just as protected as the speech itself, as when a newspaper decides which stories to print on the front page and which letters to the editor to publish. That is no less true for social media platforms. In fact, the term “content-recommendation algorithm” itself points to its expressive nature. Recommending something is a message — “I think you would find this interesting.”

The Moody Court also acknowledged the expressive nature of arranging online content (emphasis added): “Deciding on the third-party speech that will be included in or excluded from a compilation — and then organizing and presenting the included items — is expressive activity of its own.” Similarly, while dismissing exactly the kind of case the Algorithm Accountability Act would enable, the U.S. Court of Appeals for the Fourth Circuit held this past February: “Facebook’s decision[s] to recommend certain third-party content to specific users . . . are traditional editorial functions of publishers, notwithstanding the various methods they use in performing” them.

So the First Amendment is at least implicated when Congress institutes “accountability” for a platform’s arrangement and presentation of user-generated content, unlike with pharmaceutical safety regulations. But does it prohibit Congress from imposing the kind of liability the Algorithm Accountability Act creates?

Yes. Two well-established principles explain why.

First: As the Supreme Court has repeatedly made clear, imposing civil liability for protected speech raises serious First Amendment concerns.

Second: Except for the exceedingly narrow category of incitement — where the speaker intended to spur imminent unlawful action by saying something that was likely to cause such action — the First Amendment demands that we hold the wrongdoer accountable for their own conduct, not the people whose words they may have encountered along the way.

The U.S. Court of Appeals for the Fifth Circuit concisely explained why these principles preclude liability for “negligently” conveying “harmful” ideas:

If the shield of the first amendment can be eliminated by providing after publication that an article discussing a dangerous idea negligently helped bring about a real injury simply because the idea can be identified as ‘bad,’ all free speech becomes threatened.

In other words, faced with a broad, unmeetable duty to anticipate and prevent ideas from causing harm, media would be chilled into publishing, broadcasting, or distributing only the safest and most anodyne material to avoid the risk of unpredictable liability.

For this reason, courts have — for nearly a century — steadfastly refused to impose a duty of care to prevent harms from speech. A few noteworthy examples are illustrative:

• Dismissing a lawsuit alleging that CBS’ television programming desensitized a child to violence and led him to shoot and kill his elderly neighbor, one federal court wrote of the duty of care sought by the plaintiffs:

The impositions pregnant in such a standard are awesome to consider . . . Indeed, it is implicit in the plaintiffs’ demand for a new duty standard, that such a claim should exist for an untoward reaction on the part of any ‘susceptible’ person. The imposition of such a generally undefined and undefinable duty would be an unconstitutional exercise by this Court in any event.

• In a case brought by the victim of a gruesome attack alleging that NBC knew of studies on child violence putting them on notice that some viewers might imitate violence portrayed on screen, the court ruled:

[T]he chilling effect of permitting negligence actions for a television broadcast is obvious. . . . The deterrent effect of subjecting [them] to negligence liability because of their programming choices would lead to self-censorship which would dampen the vigor and limit the variety of public debate.

• Affirming dismissal of a lawsuit alleging that Ozzy Osbourne’s “Suicide Solution” caused a minor to kill himself, the court noted the profound chilling effect such liability would cause:

[I]t is simply not acceptable to a free and democratic society to impose a duty upon performing artists to limit and restrict the dissemination of ideas in artistic speech which may adversely affect emotionally troubled individuals. Such a burden would quickly have the effect of reducing and limiting artistic expression to only the broadest standard of taste and acceptance and the lowest level of offense, provocation and controversy.

• When the family of a teacher killed in a school shooting sued makers and distributors of violent video games and movies, the court rejected the premise of the suit:

Given the First Amendment values at stake, the magnitude of the burden that Plaintiffs seek to impose on the Video Game and Movie Defendants is daunting. Furthermore, the practical consequences of such liability are unworkable. Plaintiffs would essentially obligate these Defendants, indeed all speakers, to anticipate and prevent the idiosyncratic, violent reactions of unidentified, vulnerable individuals to their creative works.

In his op-ed, Sen. Curtis wrote, “The problem isn’t what users say, but how algorithms shape and weaponize it.” But the “problem” this bill seeks to remedy very much is what users say. A content recommendation algorithm in isolation can’t cause any harm; it’s the recommendation of certain kinds of content (e.g., radicalizing, polarizing, etc.) that the bill seeks to stymie.

And that content is overwhelmingly protected by the First Amendment, regardless of whether the posts might, individually or in the aggregate, cause an individual to commit violence. When the City of Indianapolis created remedies for people who viewed pornography, the U.S. Court of Appeals for the Seventh Circuit rejected the municipality’s justification that pornography “perpetuate[s] subordination” and leads to cognizable societal and personal harms:

[T]his simply demonstrates the power of pornography as speech. All of these unhappy effects depend on mental intermediation. Pornography affects how people see the world, their fellows, and social relations. If pornography is what pornography does, so is other speech.

[ . . . ]

Racial bigotry, anti-semitism, violence on television, reporters’ biases — these and many more influence the culture and shape our socialization. None is directly answerable by more speech, unless that speech too finds its place in the popular culture. Yet all is protected as speech, however insidious. Any other answer leaves the government in control of all of the institutions of culture, the great censor and director of which thoughts are good for us.

And that’s why the Algorithm Accountability Act also threatens users’ expressive rights. There’s simply no reliable way to predict whether any given post might, somewhere down the line, factor into someone else’s independent decision to commit violence — especially at the scale of modern social media. Faced with liability for guessing wrong, platforms will effectively have two realistic choices: aggressively re-engineer their algorithms to bury anything that could possibly be deemed divisive (and therefore risky), or — far more likely — simply ban all such content entirely. Either road leads to the same place: a shrunken public square where whole neighborhoods of protected speech have been bulldozed.

“What a State may not constitutionally bring about by means of a criminal statute,” the Supreme Court famously wrote in New York Times v. Sullivan, “is likewise beyond the reach of its civil law.” Forcing social media platforms to do the dirty work of censorship on pain of expensive litigation and expansive liability is no less offensive to the First Amendment than a direct government speech regulation.

Political violence is a real and pressing problem. But history has already taught us that trying to scrub away every potential downstream harm of speech is a dead end. And a system of free speech requires us to abstain from the temptation of trying in the first place.

For many people, artificial intelligence chatbots make daily life more efficient. AI can manage calendars, compose messages, and provide quick answers to all kinds of questions. People interact with AI chatbots to share thoughts, test ideas, and explore language. This technology, in various ways, is playing a larger and larger role in how we think, work, and express ourselves.

But not all the news is good, and some people want to use the law to crack down on AI.

Recent news reports describe a wave of lawsuits alleging that OpenAI’s generative AI chatbot, ChatGPT, caused adult users psychological distress. The filings reportedly seek monetary damages for people who conversed at length with a chatbot’s simulated persona and reported experiencing delusions and emotional trauma. In one reported case, a man became convinced that ChatGPT was sentient and later took his own life.

These situations are tragic and call for genuine compassion. Unfortunately, if these lawsuits succeed, they’ll effectively impose an unworkable expectation on anyone creating a chatbot to scrub anything that could trigger its most vulnerable users. Everyone, even fully capable adults, would be effectively treated as if they are on suicide watch. That’s a standard that would chill open discourse.

Like the printing press, the telegraph, and the internet before it, artificial intelligence is an expressive tool. A prompt, an instruction, or even a casual question reflects a user’s intent and expressive choice. A constant across its many uses is human agency — because it is ultimately a person that ends up deciding what to ask, what responses to keep, what results to share, and how to use the material it develops. Just like the communicative technologies of the past, AI has the potential to amplify human speech rather than replace it, bringing more storytellers, perspectives, and critiques with it.

Every new expressive medium in its time has faced public scrutiny and renewed calls for government intervention. After the famous 1938 Orson Welles’ “War of the Worlds” radio broadcast about a fictional alien invasion, for example, the Federal Communications Commission received hundreds of complaints urging the government to step in. Many letters expressed fear that this technology can deceive and destabilize people. Despite the panic, neither the broadcaster nor Welles, who went on to cinematic fame, faced any formal consequences. As time went on, the dire predictions never materialized.

Early panic rarely aligns with long-term reality. Much of what once seemed threatening eventually found its place in civic life, revolutionizing our ability to communicate and connect. This includes radio dramas, comic books, TV, and the early web.

The attorneys filing lawsuits against these AI companies argue that AI is a product, and if a product predictably causes harm, safeguards are expected, even for adults. But when the “product” is speech, that expectation meets real constitutional limits. Even when harm seemed foreseeable, courts have long refused to hold speakers liable for the psychological effects of their speech on people that choose to engage with it. For example, composing rap lyrics or televising reports of violence can’t get you sued for the effects of listening or viewing them, even if they trigger people to act out.

This principle is necessary to protect free expression. Penalizing people for the emotional or psychological impact of their speech invites the government to police the ideas, too. Recent developments in the UK shows how this can play out. Under laws that criminalize speech causing “alarm or distress,” people in England and Wales can be fined, aggressively prosecuted, or both, based entirely on the state’s claimed authority to measure the emotional “impact” of what was said. That’s not a model we should import.

A legal framework worthy of a free society should reflect confidence in adults’ ability to pursue knowledge without government intrusion, and this includes the use of AI tools. Extending child-safety laws or similar liability standards to adult conversations with AI would erode that freedom.

The same constitutional protections apply when adults interact with speech, even speech generated by AI. That’s because the First Amendment ensures that we meet challenging, misleading, or even false ideas with more speech rather than censorship. More education and debate are the best means to preserve adults’ ability to judge ideas for themselves. It also prevents the state from deciding which messages are too dangerous for people to hear — a power that, if granted, can and will almost certainly be abused and misused. This is the same principle that secures Americans’ right to read subversive books, hear controversial figures speak, and engage with ideas that offend others.

Regulating adult conversations with AI blurs the line between a government that serves its citizens and one that supervises them. Adulthood presumes the capacity for judgment, including the freedom to err. Being mistaken or misguided is all part of what it means to think and speak for oneself.

At FIRE, we see this dynamic play out daily on college campuses. These institutions of higher education are meant to prepare young adults for citizenship and self-governance, but instead they often treat students as if discomfort and disagreement are radioactive. Speech codes and restrictions on protests, justified as shields against harm, teach dependence on authority and distrust of one’s own resilience. That same impulse is now being echoed in calls for AI chatbot regulation.

Yes, words can do harm, even in adulthood. Still, not every harm can be addressed in court or by lawmakers, especially not if it means restricting free expression. Adults are neither impervious to nor helpless before AI’s influence on their lives and minds, but treating them like minors is not the solution.

Editor’s note: On October 13, Gov. Newsom vetoed SB 771, which would have had a profound chilling effect on online expression. We are grateful that the governor chose to protect free speech instead of setting California on the path to a losing battle against the First Amendment and Section 230.

This week, FIRE wrote to California Governor Gavin Newsom, urging him to veto SB 771, a bill that would allow users and government enforcers to sue large social media platforms for enormous sums if their algorithms relay user-generated content that contributes to violation of certain civil rights laws.

Subscribe now

Obviously, platforms are going to have a difficult time knowing if any given post might later be alleged to have violated a civil rights law. So to avoid the risk of huge penalties, they will simply suppress any content (and user) that is hateful or controversial — even when it is fully protected by the First Amendment.

And that’s exactly what the California legislature wants. In its bill analysis, the staff of the Senate Judiciary Committee chair made clear that their goal was not just to target unlawful speech, but to make platforms wary of hosting “hate speech” more generally:

This cause of action is intended to impose meaningful consequences on social media platforms that continue to push hate speech . . . to provide a meaningful incentive for social media platforms to pay more attention to hate speech . . . and to be more diligent about not serving such content.

Supporters have tried to evade SB 771’s First Amendment and Section 230 concerns, largely by obfuscating what the bill actually does. To hear them tell it, SB 711 doesn’t create any new liability, it just holds social media companies responsible if their algorithms aid and abet a violation of civil rights law, which is already illegal.

But if you look just a little bit closer, that explanation doesn’t quite hold up. To understand why, it’s important to clarify what “aiding and abetting” liability is. Fortunately, the Supreme Court explained this just recently — and in a case also about social media algorithms to boot.

In Twitter v. Taamneh, the plaintiffs claimed that social media platforms had aided and abetted acts of terrorism by algorithmically arranging, promoting, and connecting users to ISIS content, and by failing to prevent ISIS from using their services after being made aware of the unlawful use.

The Supreme Court ruled that they had not successfully made out a claim. Because aiding and abetting requires not just awareness of the wrongful goals, but also a “conscious intent to participate in, and actively further, the specific wrongful act.” All the social media platforms had done was create a communications infrastructure, which treated ISIS content just like any other content — and that is not enough.

California law also requires knowledge, intent, and active assistance to be liable for aiding. But nobody really thinks the platforms have designed their algorithms to facilitate civil rights violations. So SB 771 has a problem. Under the existing standard, it’s never going to do anything, which is obviously not what its supporters intend. Therefore, they hope to create a new form of liability — recklessly aiding and abetting — for when platforms know there’s a serious risk of harm and choose to ignore it.

But wait, there’s more.

SB 771 also says that, by law, platforms are considered to have actual knowledge of how their algorithms interact with every user, including why every single piece of content will or will not be shown to them. This is just another way of saying that every platform knows there’s a chance users will be exposed to harmful content. All that’s left is for users to show that a platform consciously ignored that risk.

That will be trivially easy. Here’s the argument: the platform knew of the risk and still deployed the algorithm instead of trying to make it “safer.”

Soon, social media platforms will be liable solely for using an “unsafe” algorithm, even if they were entirely unaware of the offending content, let alone have any reason to think it’s unlawful.

But the First Amendment requires that any liability for distributing speech must require the distributor to have knowledge of the expression’s nature and character. Otherwise, nobody would be able to distribute expression they haven’t inspected, which would “would tend to restrict the public’s access to [expression] the State could not constitutionally suppress directly.” Unfortunately for California, the very goal they want SB 771 to accomplish is what makes it unconstitutional.

And this liability is not restricted to content recommendation algorithms (though it would still be unconstitutional if it were). SB 771 doesn’t define “algorithm” beyond the function of “relay[ing] content to users.” But every piece of content on social media, whether in a chronological or recommendation-based feed, is displayed to users using an algorithm. So SB 771 will impose liability every time any piece of content is shown on social media to any user.

This is where Section 230 also has something to say. One of the most consequential laws governing the internet, Section 230 states, “No provider or user of an interactive computer service shall be treated as the publisher or speaker of any information provided by another information content provider,” and prohibits states from imposing any liability inconsistent with it. In other words, the creator of the unlawful content is responsible for it, not the service they used to do so. Section 230 has been critical to the internet’s speech-enabling character. Without it, hosting the speech of others at any meaningful scale would be far too risky.

SB 771 tries to make an end-run around Section 230 by providing that “deploying an algorithm that relays content to users may be considered to be an act of the platform independent from the message of the content relayed.” In other words, California is trying to redefine the liability: “we’re not treating you as the publisher of that speech, we’re just holding you liable for what your algorithm does.”

But there can be no liability without the content relayed by the algorithm. By itself, the algorithm does not cause any harm recognized by law. It’s the user-generated content that causes the ostensible civil rights violation.

And that’s not to mention the fact that because all social media content is relayed by algorithm, it would effectively nullify Section 230 by imposing liability on all content. California cannot evade federal law by waving a magic wand and declaring the thing Section 230 protects to be something else.

Newsom has until October 13 to make a decision. If signed, the law takes effect on Jan. 1, 2027, and in the interim, other states will likely follow suit. The result will be a less free Internet, and less free speech — until the courts inevitably strike down SB 771 after costly, wasteful litigation. Newsom must not let it come to that. The best time to avoid violating the First Amendment is now.

The second best time is also now.

Edan Kauer is a former FIRE summer intern and a sophomore at Georgetown University.

Elliston Berry was just 14 years old when a male classmate at Aledo High in North Texas used AI to create fake nudes of her based on images he took from her social media. He then did the same to seven other girls at the school and shared the images on Snapchat.

Subscribe now

Now, two years later, Berry and her classmates are the inspiration for Senator Ted Cruz’s Take It Down Act (TIDA), a recently enacted law which gives social media platforms 48 hours to remove “revenge porn” once reported. The bill considers any non-consensual intimate imagery (NCII), including AI deepfakes, to fall under this category. But despite the law’s noble intentions, its dangerously vague wording is a threat to free speech.

This law, which covers both adults and minors, makes it illegal to publish an image of an identifiable minor that meets the definition of “intimate visual depiction,” which is defined as certain explicit nudity or sexual conduct, with intent to “arouse or gratify the sexual desire of any person” or “abuse, humiliate, harass, or degrade the minor.”

That may sound like a no-brainer, but deciding what content this text actually covers, including what counts as “arousing,” “humiliating,” or “degrading” is highly subjective. This law risks chilling protected digital expression, prompting social media platforms to censor harmless content like a family beach photo, sports team picture, or images of injuries or scars to avoid legal penalties or respond to bad-faith reports.

Civil liberties groups such as the Electronic Frontier Foundation (EFF) have noted that the language of the law itself raises censorship concerns because it’s vague and therefore easily exploited:

TAKE IT DOWN creates a far broader internet censorship regime than the Digital Millennium Copyright Act (DMCA), which has been widely abused to censor legitimate speech. But at least the DMCA has an anti-abuse provision and protects services from copyright claims should they comply. This bill contains none of those minimal speech protections and essentially greenlights misuse of its takedown regime … Congress should focus on enforcing and improving these existing protections, rather than opting for a broad takedown regime that is bound to be abused. Private platforms can play a part as well, improving reporting and evidence collection systems.

Nor does the law cover the possibility of people filing bad-faith reports.

In the 2002 case Ashcroft v. Free Speech Coalition, the Court said the language of the Child Pornography Protection Act (CPPA) was so broad that it could have been used to censor protected speech. Congress passed the CPPA to combat the circulation of computer-generated child pornography, but as Justice Anthony Kennedy explained in the majority opinion, the language of the CPPA could be used to censor material that seems to depict child pornography without actually doing so.

Also in 2002, the Supreme Court heard the case Ashcroft v. ACLU, which came about after Congress passed the Child Online Protection Act (COPA) to prevent minors from accessing adult content online. But again, due to the broad language of the bill, the Court found this law would restrict adults who are within their First Amendment rights to access mature content.

Share

As with the Take It Down Act, here too were laws created to protect children from sexual exploitation online, yet established using vague and overly broad standards that threaten protected speech.

But unfortunately, stories like the one at Aledo High are becoming more common as AI becomes more accessible. Last year, boys at Westfield High School in New Jersey used AI to circulate fake nudes of Francesca Mani, who is 14 years old, and other girls in her class. But Westfield High administrators were caught off guard as they had never experienced this type of incident. Although the Westfield police were notified and the perpetrators were suspended for up to 2 days, parents criticized the school for their weak response.

A year later, the school district developed a comprehensive AI policy and amended their bullying policy to cover harassment carried out through “electronic communication” which includes “the use of electronic means to harass, intimidate, or bully including the use of artificial intelligence "AI" technology.” What’s true for Westfield High is true for America — existing laws are often more than adequate to deal with emerging tech issues. By classifying AI material under electronic communication as a category of bullying, Westfield High demonstrates that the creation of new AI policies are redundant. On a national scale, the same can be said for classifying and prosecuting instances of child abuse online.

While we must acknowledge that online exploitation is a very real issue, we cannot solve the problem at the expense of other liberties. Once we grant the government the power to silence the voices we find distasteful, we open the door to censorship. Though it is essential to address the very real harms of emerging AI technology, we must also keep our First Amendment rights intact.

Social media platforms are having a First Amendment moment. From content moderation to so-called “addictive” features such as unlimited scrolling, lawmakers are putting bipartisan pressure on these companies to change their policies and even the fundamental designs of their platforms. But to do so, they have to first define what a social media platform is.

Subscribe now

That task has fallen to lawmakers and judges across the country, all struggling to arrive at a definition. However, their reasons for doing so and the approaches they are taking reflect the peculiarities of their respective branches: Lawmakers need a definition they can use to craft workable policies, while judges seek to draw connections between today’s social media cases and established precedent from cases involving past communicative technologies. But the resulting vague laws and inconsistent legal analogies show the severity of these definitional difficulties.

For example, Ohio’s law defines a social media company as an online website, service, or product that allows users to “interact socially” by building a profile, developing a list of contacts, and creating or sharing content.

Despite its relative clarity compared to many other states’ laws, even this definition has some glaring holes. For one, how are courts to decide what it means to “interact socially” on any given platform? Of course, direct messaging on Instagram or reposts on X clearly qualify, but what about building a collaborative playlist on Spotify or saving someone’s Pinterest post to your own board? These examples may seem like a stretch, but both involve people interacting in a way that communicates some message.

The Ohio law also makes exceptions for comment sections under “content posted by an established and widely recognized media outlet.” But the law does not define what makes a media outlet “established and widely recognized.” This vagueness opens the door to partisan abuse when enforcing the law. A liberal attorney general might attempt to enforce the law against Breitbart News, while a conservative one could punish The Huffington Post. And by excluding some topics but not others, the law again reveals itself to be a content-based (and therefore likely unconstitutional) regulation.

When trying to curtail platforms' content moderation policies, Texas and Florida took their own stabs at defining them. They, too, missed the mark: Florida Senate Bill 7072 is incredibly broad, lumping together social media, search engines, and any "internet platform" that meets the revenue and user thresholds. Texas House Bill 20 gives a leaner definition, but still clumsily targets public, account-based platforms that are designed to let users communicate. It explicitly excludes email and news outlets, revealing itself as an attempt to carve social media out of otherwise-protected online speech.

On the judicial side, the Texas and Florida laws were challenged by NetChoice, a trade group representing social media companies. In both cases, judges sought to understand what social media platforms are by looking at what they do — and how it fits with past case law. In Moody v. NetChoice, the U.S. Court of Appeals for the Eleventh Circuit ruled against Florida, reasoning that social media platforms behave like newspaper publishers or parade organizers, whose curation of speech in their products is protected by the First Amendment.

In NetChoice v. Paxton, the Fifth Circuit took a different approach by likening platforms not to assemblers and curators of speech but rather hosts or conduits of speech. According to this definition, moderation decisions are "not speech at all" and do not enjoy First Amendment protection. The Supreme Court heard both cases together and immediately sent them back to the lower courts for further consideration on technical grounds — but in doing so it signaled that a supermajority of the Court agreed that the Fifth Circuit’s analysis was wrong, and that social media platforms’ content policies and decisions are protected expression the same as they would be for a newspaper.

But just because lawmakers and judges take different approaches to defining social media platforms doesn't mean their roles can’t synergize to positive ends. Look no further than Section 230 of the Communications Decency Act, which shields online services from liability for third-party content published on their sites. Section 230 predates modern social media and thus doesn't offer a definition for platforms, but it helped define the internet — and social media — as we know them today.

Soon after Section 230 passed, courts were asked to clarify the law’s application. In Zeran v. America Online, the Fourth Circuit interpreted Section 230 as saying that platforms can’t be sued for carrying out “a publisher’s traditional editorial functions.” Accordingly, Section 230 confers an additional layer of protection on social media platforms’ content decisions: The First Amendment protects platforms' editorial decisions about which content to display, and Section 230 shields them from liability for the third-party content that is displayed and their decisions about whether or how to display it. The story of Section 230 demonstrates how the courts can apply well-written laws to protect online speech.

Share

This process has yet to result in a definition for social media platforms. However, a definition could be beside the point: Section 230 shields online platforms from liability for carrying third-party content, and just because social media is a relatively new phenomenon does not mean the First Amendment does not apply.

A big reason lawmakers seem to get tripped up in answering this question is because in the past, different forms of media had singular functions: reading newspapers, listening to the radio, or watching and listening to the television — this was part of the basis for how media was regulated. With today's technological advances, different forms of communication have converged and the old regulatory model has been complicated. Attempting to treat a complex platform like Facebook as a telegram company is not only unproductive, it discourages a vibrant online marketplace of ideas.

Instagram publishes third-party content like a newspaper, enables direct communication like a phone company, and makes decisions about which speech it includes and excludes like a parade organizer. Courts have made it clear that as new speech mediums emerge, whether in film, online, or in video games, the First Amendment's governing principles apply with the same force.

Current efforts to define social media are motivated by a desire to control it and, by extension, speech. So far, courts have mostly shown respect for the First Amendment and struck down bad laws — although the Supreme Court's recent decision in Free Speech Coalition v. Paxton could be a warning that judges are starting to turn their backs on protecting online speech.

Social media companies are the latest communicative technology to be caught in the crosshairs of frightened legislatures. When drafting laws to define social media, lawmakers should remember that while the medium has changed, the speech at issue remains protected. And when judges read those laws, they should maintain the respect that courts have shown to the First Amendment as applied to new technologies.

This essay was originally published by The Dallas Express on July 21, 2025.

Don’t want to publish an opinion on your blog you disagree with? Too bad, the government forces you to publish it. Criticize the mayor? Go to jail — and good luck trying to sue the mayor for violating your First Amendment rights. Want to access online content legal for adults without jeopardizing your privacy and reputation? Think again, your state legislature demands you reveal your identity first.

Subscribe now

That’s not the America I know. Nor is it one a robust First Amendment would ever allow. But those constitutional threats came up before the Supreme Court during its past two terms, thanks to state legislatures and government officials who would prefer to play thought police instead of obey the First Amendment’s commands.

For the most part, the Supreme Court did as it has for decades: It upheld the First Amendment as a mighty check on government intrusion into our thoughts, on public debate, and in the search for truth. Still, a couple of the Court’s decisions this year broke from that trend, causing First Amendment defenders everywhere to raise their collective eyebrows.

Americans’ pilgrimage to the online world has placed free speech on the internet at the forefront of the Court’s recent First Amendment decisions. And that includes a case involving our Lone Star State. Last July, in Moody v. Netchoice, the Supreme Court considered Texas and Florida laws that tried to dictate how social media companies decide what political and ideological content to allow, and how to present it.

Before the Supreme Court sent the cases back to the appeals courts for another look, it made one thing clear: The First Amendment bars the government from telling social media platforms what they can and can’t publish, just as it bars the government from telling newspaper editors what they can or can’t print. As Justice Kagan remarked, “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.”

In another pair of rulings, the Supreme Court considered two instances of local officials blocking citizens from commenting on the officials’ social media pages about the government’s performance. Although the Court didn’t rule in favor of either party, it confirmed the First Amendment limits officials’ power to block Americans from commenting on social media pages an official uses to conduct government business.

With these social media decisions, the Court underscored that core First Amendment principles apply just as strongly in the digital age, helping to secure free speech online from government overreach.

On the other hand, this most recent Supreme Court term found the Court twice deferring to governmental regulations of online expression. First, in TikTok v. Garland, the Court held the federal government’s effective ban on the popular social media app TikTok does not violate the First Amendment, even as half the United States uses the platform to speak and to receive information. The Court largely yielded to the government’s asserted concerns over national security, despite, as many pointed out, Congress’s failure to provide enough evidence showing TikTok poses a national security threat. In fact, the administration’s continued unwillingness to enforce the ban punctuates how suspect Congress’s national security concerns were.

Share

And in another decision, Free Speech Coalition v. Paxton, the Court upheld Texas’s law requiring adults to verify their age before accessing websites with sexually explicit material that is, while legal for adults, “obscene for minors.” Breaking from decades-old prior precedent that invalidated similar laws, the Court held the First Amendment doesn’t protect adults “accessing material obscene to minors” until they verify their age — even though that material enjoys full First Amendment protection.

While the Court noted the accepted practice of ID checks for sexually explicit material in the physical world, it all but skirted the unique and serious privacy implications digital ID checks impose, in a time where Americans suffer harm from regular data breaches. Those privacy scares will chill many adults from seeking speech the First Amendment protects for their use.

For all that, both the TikTok and Free Speech Coalition decisions are narrow ones. So on the whole, they should not undermine the broad protections for internet speech the Supreme Court has confirmed in prior terms. Still, any time the government censors speech on shaky evidence about national security concerns, or tries to burden adult access to protected speech in the name of childproofing the internet, it raises a First Amendment red flag that should concern us all.

Outside of the digital world, the Supreme Court handed down several decisions over the last two terms vindicating the First Amendment as a vital check on overzealous government officials abusing their power to silence opinions they don’t like. For instance, in NRA v. Vullo, the Supreme Court held officials in New York violated the First Amendment when they coerced financial services and insurers and underwriters to sever ties with the National Rifle Association — all because the state disagreed with the NRA’s constitutionally protected advocacy. That’s hugely important for freedom of speech, affirming that the government can’t strong-arm third-parties into silencing Americans because of the views they express.

Another decision centered on a San Antonio-area woman was a good step towards ensuring government critics have a robust remedy when officials wrongfully arrest them. In Gonzalez v. Trevino, the Supreme Court clarified that Sylvia Gonzalez, a long-time government critic, could sue the local officials who singled her out for arrest after she called for the city manager’s removal.

And just this May, the Supreme Court halted the Maine Legislature’s denial of State Representative Laurel Libby’s right to vote just because the legislature’s political majority took offense to the lawmaker’s First Amendment-protected social media post about a transgender athlete participating in a high school event. Not only did the majority’s act infringe Rep. Libby’s First Amendment right to comment on public issues, it also deprived her constituents of the representation our republican form of government guarantees.

At its core, the Court’s ruling in Libby v. Fecteau underscored a vital constitutional principle: Political majorities cannot censor and exclude others from the democratic process based on the views they express. That’s a first principle worth upholding, no matter where a speaker falls on the ideological spectrum.

And we should all be glad the Supreme Court upheld that principle here, in a time where protecting the uniquely American freedoms to dissent and voice our opinions without fear of the government’s strong hand is as important as ever.

This essay was originally published in 24sight’s The Vox Populi section on June 24, 2025.

As artificial intelligence plays a growing role in political discourse, lawmakers across the country have rushed to propose new regulations over fears that misinformation will proliferate with the new technology. But new polling suggests these efforts may not fully reflect public sentiment and serve as a red flag for lawmakers when their state legislatures reconvene in the fall.

Subscribe now

A recent national survey conducted by Morning Consult for the Foundation for Individual Rights and Expression finds that American voters strongly support prioritizing free speech when crafting AI regulations, even amid growing concerns about AI’s impact.

Sixty percent of voters say AI-generated content poses a greater threat to elections than government regulation of it. Yet when pressed to choose between stopping deceptive content and protecting free speech, voters side with free speech, 47% to 37%. That support cuts across political lines.

These numbers also tell a complicated story: Americans are uneasy about AI, but they’re more concerned about the government using AI regulation as a tool to silence dissent, just like many Americans feel about the government regulating content on social media. Our most recent survey shows a striking 81% worry that rules governing election-related AI content could be misused to suppress criticism of elected officials. And over half fear that making it a crime to publish altered political content could chill legitimate political commentary.

These aren’t abstract fears. Across the country, lawmakers have introduced, and even enacted, bills that would target the mere sharing of AI-generated political content, no matter the context or intent. In Texas and Vermont, for example, proposed legislation cast such a wide net that distributing satire, parody, criticism, or even memes would have been banned or otherwise ensnared in regulation. Ordinary citizens, not just political campaigns, would face penalties for posting altered images of politicians online.

Share

Americans have the right to speak, joke, criticize, and comment freely, regardless of whether they use AI as an expressive tool in doing so. When lawmakers write vague or sweeping rules about what people can say about candidates, they silence the very public discourse that elections foster.

Indeed, 28% of voters say government regulation of AI-generated or AI-altered content would make them less likely to share content. That’s not just a statistic, it’s a warning sign. Lawmakers risk silencing voters when their voices matter most. And the effect is even greater among young people, who are significantly more likely to engage with and create AI-generated content. When nearly a third of voters, especially the next generation of political voices, are deterred from participating in public discourse, we’re not just regulating technology — we’re shrinking the space for political engagement.

Not every datapoint in the polling breaks in favor of free speech. Protecting speech commands broad support, yet many voters also favor checks on misinformation. In the same FIRE survey, while 77% of voters think preserving the right to freedom of speech should be the government's main priority when making laws that govern the use of AI, 74% of voters believe it’s more important to protect people from misinformation than it is to protect free speech.

Even so, the poll suggests many voters want any effort to curb misinformation to have firm safeguards for open debate. Many bills on the table this year definitely missed that mark.

Technologies evolve. The principles of the First Amendment do not. In our system of government, the answer to bad speech isn’t censorship. It’s more speech.

John Coleman is legislative counsel and Sean Stevens is chief research advisor for the Foundation for Individual Rights and Expression (FIRE).

Subscribe now