Artificial intelligence is a powerful tool for free expression. And it keeps improving. As the saying goes, it’s the worst it will ever be. People use it to receive ever-smarter speech and to create ever-better speech of their own. AI is now an important part of the marketplace of ideas—a market that, in America, is not for the state to control. The more useful AI becomes, the more worried you should be about letting the government dictate what AI can and cannot say. Like the powerful speech tools that came before it—the printing press, the camera, the internet—AI warrants full First Amendment protection.
The right question is not whether AI outputs are protected, but how. The key is to insist that they receive the same First Amendment protection as other expressive technologies. No more, to be sure; but also no less. Striking that balance may require some new rules, crafted with free-speech fundamentals firmly in view. This might sound like a daunting task at a time when even many AI researchers are losing their heads over AI’s purported existential risk. But a little common sense can carry us a long way.
For a start, we can draw on the venerable First Amendment distinction between speech and conduct. When someone uses AI to obtain information—to learn and to think—she is exercising her constitutional rights as a listener. When she uses AI to help her write an essay or create a piece of art, she is exercising her constitutional rights as a speaker. The government can no more restrict these uses of AI than it can restrict the use of an encyclopedia or a thesaurus, a paintbrush or a typewriter. The First Amendment has nothing to say, however, about using AI to hack a website or rob a bank. These acts are illegal without AI, and they remain illegal with it. AI is not entitled to special constitutional privileges.
In narrow circumstances, speech alone can break the law, and here things get trickier. Several categories of speech unprotected by the First Amendment, such as incitement to violence and criminal conspiracy, require intent. You must have meant to bring about the violence or commit the crime. A chatbot does not “intend” to do anything, even if it talks as though it does. So does no one answer for what a chatbot says?
One approach is to look to the intent of the firm that built the model. Did the firm want the AI to incite or conspire? Almost certainly not. This would yield very broad protection indeed—too broad, a lot of people would say. Suppose a chatbot actively helps a user plan a mass shooting. The user announces his plan, and the AI goes right along. A firm that argues, in its defense, that its AI merely mimics intent can expect an offended public and a skeptical judge. A human who said the same things could face charges as an accomplice. Why treat the firm any differently? This looks, many will think, like extra speech protection, rather than equal speech protection, for artificial intelligence.
It bears mention that there is a serious argument the other way. We’ve long had works that express an “intent” belonging to no one in particular. They’re called novels, poems, and songs. In The Sorrows of Young Werther (1774), the protagonist intends to tell you that suicide is an appropriate response to unrequited love. It does not follow that the author, Johann Wolfgang von Goethe, intended to tell you the same thing. Artists take us to dark places as an exercise of imagination, and we don’t punish them for doing the job well. On the contrary, we celebrate great artists for creating art that feels authentic. Our safety-first culture is unlikely to leave it entirely to users to grasp that AIs don’t “intend” things. But such reliance on personal responsibility would be a perfectly civilized approach. The alternative—barring all speech that might send a handful of vulnerable people in a bad direction—carries dangers of its own. Today, ChatGPT. Tomorrow, Goethe. Then Marx and Nietzsche. Then anything offensive.
At the other extreme, we might hold AI firms strictly liable for any functional intent their models display. If a model talks as though it wants to help a user commit a crime, the firm is liable for aiding and abetting. But that won’t do. For the firm, such a regime might be even worse than having its model’s outputs treated as no speech at all. Ascribe the model’s every word, intent and all, to the firm, and the firm could face criminal liability even when it has exercised reasonable care. For the public, meanwhile, the chill on AI outputs would be almost as bad as consigning the models to government censorship. The only safe course for a firm would be to constrict its model until it refused to discuss any remotely controversial topic.
The most workable middle ground is likely some kind of principal-agent framework. Think of the relationship between a company and its employees. A company is responsible for making sure its employees behave themselves, but the employees’ intentions are not automatically the company’s own. If an employee goes rogue and commits a crime on the clock, the company is usually not criminally liable itself. In most cases, though, the company will face civil liability for the harm an employee causes within the scope of employment. If an employee incites violence, the company is not treated as though it incited violence, but it still may have to pay for the harm the employee caused. This gives the company good reason to hire carefully, to train its employees well, to set clear rules against misconduct, and so on.
Even here, we should keep squarely in mind that it’s speech we’re dealing with. AI models host billions of conversations. An AI firm knows that in this vast sea of discourse, a small fraction of conversations will go sideways. It is foreseeable that a few vulnerable users will have bad experiences. Deeply troubled people will do the things that deeply troubled people do, and they will sometimes do them with some encouragement—or, more often, ineffective discouragement—from a chatbot. If an AI firm is treated as the insurer of everything its model says to many millions of users, that too will chill speech. Firms will face pressure to over-sanitize their models, making them far less useful to the vast majority of us who can use them responsibly. Your rights as a listener and as a speaker should not be shackled to the needs of the most fragile person imaginable. Yet that’s where we’ll end up if we hold AI firms strictly liable for the handful of cases in which their guardrails give way in extreme situations, with users who, AI use or no, are a danger to themselves or others. A censorship aimed at shielding them is a censorship of the entire public. The marketplace of ideas will be far the worse for it.
If an AI firm deliberately sets up its model to engage in unlawful speech, throw the book at it. That’s not the hard case. The leading AI firms don’t want their models to break the law. For those firms, a serviceable rule—one that puts AI outputs on a par with other speech—would impose liability when (1) the firm knew that its model tended to lapse into some type of unprotected speech; (2) the firm deployed the model anyway, or kept it in service, without taking adequate steps to address the problem; and (3) the model went on to engage in that speech with multiple users. In those circumstances, the firm would not be criminally liable for the model’s speech, since its executives still would not have intended to incite violence or what have you (unless separate evidence showed that they did). But the firm would very much be civilly liable for failing to stop its agent from repeatedly saying constitutionally unprotected things.
AI firms now face lawsuits over tragedies such as suicides and school shootings. At least some of these suits seek to hold an AI firm liable for outputs that would not be illegal if spoken by a human. A chatbot talks to a troubled person about guns, and the troubled person then commits violence. A chatbot talks to a troubled person about suicide—including urging against it—and the troubled person then takes their own life. The connection is likely no more than correlational: a troubled person does what troubled people do, and happens to have talked to a chatbot along the way. Yet plaintiffs’ lawyers seek to use these tragedies to convince courts to change what chatbots can say: no memory of past conversations, no first-person pronouns, nothing “emotional” or “agreeable.” What these lawyers seek—sometimes subtly, sometimes not so subtly—is to impose on AI a standard unknown to humans.
Governments are getting in on the act as well. Florida’s attorney general plans to push legislation that would hold an AI firm criminally liable if a model “participates” in a crime. No bill is out yet, but the idea seems to be to punish a company whose model simply answers questions for a criminal who never discloses his criminality. By that logic, a chatbot could not tell a hunter how his rifle’s safety works, since the next person to ask might be a killer. (If anything, this understates the problem. One suit faults a chatbot for telling a shooter when a student union is busy—a question someone simply trying to avoid the lunch rush might ask.) These lawyers and officials want AI firms to pay for outputs that would be protected expression if spoken by you or me. They are trying to produce precisely the chilling effect the First Amendment is meant to foreclose.
At bottom, they aren’t taking the First Amendment seriously. They’re making the same error as so many people who say, “I support free speech, but . . . .” They want the First Amendment to protect speech that is “good,” by their lights, and to abandon speech they deem “bad.” What’s more, they pretend the line between the two is clear and easy to administer. In reality, the First Amendment protects a lot of messy and even “bad” speech. It generally protects lies, bigotry, hate speech, and calls for anarchy or revolution. It does so not because all this speech is “good”—though sometimes it is; today’s sedition can be tomorrow’s social justice—but because it is not for the state (or the plaintiffs’ bar) to separate “good” speech from “bad.” It is for you—a citizen, not a subject—to make these judgments for yourself. As a listener, you have a right to the full marketplace of ideas. And a speaker, for her part, has no duty to make sure her words cannot hurt your feelings or harm a fragile listener.
So we return to where we began: It is crucial that AI outputs receive full First Amendment protection. This isn’t only about protecting AI speech (though it is about that too, of course). It’s also about keeping a censorship first applied to AI from contaminating First Amendment analysis writ large. That prospect should scare you. What if your friend, child, or spouse dies by suicide or commits murder? If you’re held to the proper First Amendment standard, you have little to fear from the law, so long as you didn’t overtly encourage the act. But if you’re held to the standard many want to apply to AI, you have every reason to worry about anything you said. Your every conversation with that person is now fair game for scrutiny. Your every word is now suspect. Did you fail to catch the signs? Did you fail to push back on their designs, even their oblique suggestions, with adequate vigor? Did you fail to report them to the authorities at the first hint of trouble? Were any of your offhand remarks insufficiently careful? If you engaged with the person at all, you must expect a lawsuit and potential liability.
“If there is any fixed star in our constitutional constellation,” Justice Robert H. Jackson wrote, “it is that no official, high or petty, can prescribe what shall be orthodox in politics, nationalism, religion, or other matters of opinion.” As we navigate a revolutionary new technology, we should keep that revolutionary old idea in mind. The First Amendment has served this country well. AI does not need a new set of rules, built from scratch. It needs the rules we have long had, with minor adjustments to keep them current as technology advances. The state may not do to a chatbot what it may not do to a journalist, a librarian, or you.
Corbin K. Barthold is a senior litigation fellow at NetChoice.
