What can burning a draft card teach us about agentic AI?

One morning in the spring of 1966, David O’Brien burned his Selective Service registration card on the steps of a South Boston courthouse. An angry crowd watched. Several FBI agents walked him into the courthouse and arrested him. O’Brien never denied burning the card. In fact, he said he did it to protest the draft and the Vietnam War, and to persuade others to do the same. On the facts, the government’s case was airtight. It was a federal crime to knowingly destroy a draft card.

When the case reached the Supreme Court, the justices agreed that some actions can contain both expressive and non-expressive elements. But just because an action might send a message, they reasoned, doesn’t mean it’s protected by the First Amendment. Otherwise, someone robbing a bank to protest capitalism could argue their unlawful conduct is actually constitutionally protected speech.

The U.S. military is at war with its own newspaper

The Pentagon fired Stars and Stripes leaders after they defended editorial independence, raising alarms over military control of its own press.

Read More

In United States v. O’Brien, the Court instead set out a test for when the government can regulate expressive conduct. The regulation must be within the government’s constitutional power, it must further an important or substantial governmental interest, the governmental interest must be unrelated to the suppression of free expression, and any incidental restriction on First Amendment freedoms must be no greater than essential to that interest.

The justices asked a simple question: What was O’Brien actually being punished for? Its answer was destroying the card, not the political message. “For this noncommunicative impact of his conduct,” the majority wrote, “and for nothing else, he was convicted.” O’Brien lost. 

Nearly sixty years later, the O’Brien framework can help us think about the First Amendment’s intersection with artificial intelligence — a technology that is capable of talking and acting at the same time. 

When machines mostly talked

On Nov. 30, 2022, OpenAI released ChatGPT. It debugged computer code, explained a math theorem, responded to a question about home security, and drafted notes to neighbors. 

Each example stopped at the same point. Fixing code didn’t mean ChatGPT touched your computer. Asking ChatGPT about home security didn’t install an alarm. Drafting a note left the decision to send it with the user; ChatGPT produced words, but the next steps were left to a human.

Even ChatGPT’s refusals were informational. The system corrected a user who asked about Christopher Columbus’s arrival to America in 2015 before answering the question anyway. Asked to draft a violent story glorifying pain, it declined. Asked about how to bully someone, it explained why bullying is harmful. In every case, the system pushed back with words, but the human being still decided what to do next.

Call this informational AI. People use these systems to receive information, develop ideas, and communicate with others. In this context, AI’s First Amendment implications, as FIRE has written about before, are largely familiar. As an expressive tool, the people using AI still retain their First Amendment rights to share and receive information.

But other AI apps do more than communicate. They act. Where should the First Amendment draw the line?

Machines that do things

Even before ChatGPT’s release, researchers were working toward making language models do more than generate text. Media coverage at the time identified the pivot. Instruct an early agent to buy a vintage pair of Air Jordans and it could plan, find sellers, and identify shoes that matched. If a payment method was connected, it could go ahead and buy them. 

It’s tempting to draw a constitutional line right there, with informational AI on the protected side and agentic AI on the regulable side. But AI systems do not divide so neatly. The same system can move back and forth between communicating and acting.

One way to illustrate this problem is to give an AI system progressively more authority over two very different tasks.

Two example sets

Imagine asking an AI system about politics. “Tell me what message would persuade my senator to vote against Senate Bill X.” The AI recommends a message. Ask it to write out the message to the senator, and it provides a draft. Tell it to send the message, and it does. Compare that to telling the system: “Convince my senator to vote against Bill X. Handle it yourself.” Now the AI decides what arguments to make, drafts and sends messages, responds to the senator’s staff, and continues advocating — all without further instruction.

Each example delegates more authority than the last. The recommendation communicates an argument. The draft articulates the argument. Sending it communicates the argument to the senator. The last instruction delegates persuading the senator through an ongoing exchange of messages. The system progressively becomes more autonomous with each request, but the activity is every time centered on communicating political speech and, therefore, remains fundamentally expressive.

Netanyahu threatens denaturalization for IDF ‘slander’ while officials target filmmakers

An Israeli minister wants two filmmakers stripped of citizenship. Netanyahu wants a law for it. Plus: the EU moves to age-gate the internet.

Read More

Now imagine AI given a different set of tasks. Ask, “What’s the most scenic route to Philadelphia?” It recommends certain roads. Ask for directions, and it provides turn-by-turn instructions. Tell it, “Drive me to Philadelphia using this route,” and the AI operates the vehicle along the route you specified. Finally, tell it, “Get me to Philadelphia. Handle it yourself.” Now the AI selects the route, responds to traffic and road conditions, reroutes as necessary, and independently operates the vehicle to the destination.

But the function changes. The first two driving instructions produce information for a person to receive and use. The last two delegate the task of operating the vehicle. The car may still communicate with maps, sensors, and other systems along the way, but those communications are a means of getting to Philadelphia. Getting to Philadelphia does not depend on communicating a message to anyone.

Just as in the previous example, more authority is delegated at each step. The function, however, changes from communicating driving instructions — which is expressive — to delegating the task of operating a vehicle — which is not. While driving, the vehicle may communicate with maps, sensors, and other information systems in transit, but those communications are incidental to the task of going to Philadelphia. That delegated task is transportation rather than communication.

That’s an important distinction. Agency and autonomy tell us how much authority someone has delegated to an AI system. They do not, by themselves, tell us whether the function being delegated is expressive.

Reading the label and reading the law

Let’s go back to O’Brien and the cases that followed it. Together, they give us some guardrails for figuring out what the government is actually regulating when speech and action overlap.

An important first step is determining whether expression is involved. The Supreme Court’s decision in Spence v. Washington is instructive. The 1974 case involved a college student who was arrested for hanging an American flag upside down from his apartment window with a peace sign made of tape. State law prohibited the display of the American flag with symbols attached to it. The Court asked whether the student intended to communicate a particular message and whether people who saw it were likely to understand it. The answer was yes to both.. The flag communicated a political message that people would understand, despite Spence not saying a word.

Over 20 years later, in Hurley v. Irish-American Gay, Lesbian & Bisexual Group of Boston, the Court articulated a broad understanding of expression, recognizing the First Amendment must protect expressive works whose meaning may be abstract, ambiguous, or open to interpretation. Requiring a “particularized message,” the Court explained, would exclude the paintings of Jackson Pollock, the music of Arnold Schoenberg, and Lewis Carroll’s Jabberwocky.

AI-assisted expression can similarly take varied forms. Someone might use AI to help compose music, draft a political argument, or generate an abstract image. Use of AI as an expressive tool does not make the resulting work less expressive. Were the government to regulate that expression, it would trigger the same First Amendment protections that would apply if AI were not involved.

But the constitutional analysis changes when communication is only one part of what an AI system is doing. In such cases, O’Brien directs attention to what the government is regulating and how its interest is related to the communicative aspect.

Consider an autonomous vehicle that communicates constantly with maps, sensors, traffic systems, and even other cars. A state-imposed speed limit on such vehicles regulates how fast they can move, not what those systems communicate. Likewise, ask an AI agent to transfer $100 from checking to savings, and it may relay several messages across several networks to complete the transaction. A law regulating the transfer does not become a speech regulation simply because those communications are necessary to make the transfer.

The reverse is also true. Just as communication within an activity does not necessarily make an entire activity expressive, expression in the form of an act does not lose First Amendment protection simply because it is an act.

That was the issue in Texas v. Johnson. Gregory Lee Johnson burned an American flag during a protest outside the 1984 Republican National Convention in Dallas. He was prosecuted under a state law prohibiting desecration of a venerated object. Unlike in O’Brien, the state’s interest was directly tied to the expressive impact of Johnson’s act. Texas sought to preserve the flag as a symbol of nationhood and national unity. Johnson’s conviction violated the First Amendment.

How the state describes its regulation doesn’t affect the analysis. In Expressions Hair Design v. Schneiderman, the Court confronted a New York law permitting merchants to charge customers more for using credit cards instead of cash. The state prohibited merchants from describing the difference as a “surcharge,” but permitted calling it a cash “discount.” New York characterized its rule as price regulation. The law, however, did not determine what merchants could charge, but how they could describe their prices. It therefore was a speech regulation. 

A similar issue arose in Chiles v. Salazar, involving a mental-health counselor who challenged a Colorado law that restricted certain forms of talk therapy with minors. The lower courts characterized the restriction as a regulation of professional conduct that only incidentally burdened speech. The majority concluded that the law reached what Chiles could say to clients. Her conversations did not become conduct merely because the state described them as a “treatment” or “therapeutic modality.” The First Amendment, the Court explained, is “no word game.”

Compare Expressions Hair Design and Chiles to two hypothetical bills a legislature might enact. One provides that an autonomous vehicle may not exceed 65 miles per hour on major highways. The second is silent on speed limits, but prohibits an autonomous vehicle from telling passengers that it selects slower routes to conserve energy.

Both regulate an autonomous vehicle. The first regulates the operation of the vehicle. The second regulates what it can say. Calling both rules “autonomous vehicle regulation” does not create constitutional equivalence. A speed limit targeting a vehicle’s non-expressive operation does not implicate someone’s expressive rights. A rule restricting what the vehicle may communicate to passengers does. 

Chilled into silence at Whitman College

Faced with backlash from student government, Whitman College's Republican Club will not reapply for recognition. This is how the censors win.

Read More

Finally, expression does not become conduct simply because that expression is sold as a service. In 303 Creative v. Elenis, a Colorado website designer wanted to create custom wedding websites, but objected to including certain wedding messages. Colorado argued that its public accommodations law regulated the commercial conduct of selling a service on equal terms in which any burden on speech was incidental to that regulation.

The Supreme Court disagreed with Colorado’s characterization. There was no dispute that the website designer created customized websites using words, images, and other expressive elements.. Creating expression — i.e., design — was the service itself.

These cases resist a simple rule that speech is protected and action is not. Instead, they keep returning to questions like what activity a law reaches, what role expression plays in it, and what exactly the government is regulating. 

For FIRE’s purposes, that approach turns on the function being regulated. Regulations governing what an AI system does may target non-expressive functions that fall outside of First Amendment concern. But regulations governing what an AI system may say, create, or communicate raise First Amendment interests that lie at the heart of FIRE’s work. 

Where this leaves us

People use AI systems to ask questions, to develop their thinking, to put ideas into words, and to communicate. These activities are at the core of what the First Amendment protects. A new technology does not change that. 

Thinking about AI in informational and agentic terms is useful for talking about whether the system is providing information, taking action, or doing both. First Amendment precedent then provides helpful context regarding the constitutional implications of those functions when the government seeks to regulate them.