The Listener’s First Amendment: Chatbot Output, Coverage, and the Limits of Liability

When fourteen-year-old Sewell Setzer III took his own life after months of intimate, sexualized exchanges with a companion chatbot, the firm that built it immediately reached for the First Amendment. Character Technologies argued that its users had a First Amendment right to receive what the chatbot produced, and that the precedent shielding books, films, and video games from tort liability should shield a large language model (LLM) as well. Judge Anne Conway denied the motion. She accepted that Character Technologies could assert the First Amendment rights of its users, yet she was “not prepared to hold that Character A.I.’s output is speech” at this stage of the case. [1] The company thus had standing to assert a right to receive speech the court would not yet call speech.

That tension is where this Article begins, because it shows the prevailing debate has started in the wrong place. Litigants and courts alike have framed the threshold question as one about the machine, asking whether the model is a speaker and whether what it emits counts as its speech. That framing is a mistake. The First Amendment has always been, in part, a charter for audiences, and the audience here is human. A user’s right to receive a chatbot’s output should trigger scrutiny whether or not the machine qualifies as a speaker. The real question is whether a doctrine built around listeners can be cabined before it swallows every regulation of every product that happens to emit words. It can be, though the line that does the cabining is more fragile than its proponents admit.

The State of the Litigation

Two kinds of suit are reaching the courts. In the first, the plaintiff was harmed by what a chatbot said to him, as in Garcia and the parallel suits against OpenAI and Character Technologies. [2] In the second kind, the plaintiff was harmed by what a chatbot said about him, which makes the claim defamation, as in Walters v. OpenAI.

Garcia arose from those exchanges. Character Technologies moved to dismiss on First Amendment grounds, and the court divided the question. It held that the company could assert its users’ First Amendment rights, an application of third-party standing. It then declined to take the next step, writing that the defendants had not articulated why words strung together by an LLM are speech. The refusal was tied to the posture of the motion rather than offered as a holding. [3] It let the design-defect and failure-to-warn claims proceed, treating Character.AI as a product for that purpose without holding that it is one. [4] Liability could therefore move forward on a product theory while the question of speech stayed open. The case settled in January 2026, so no appellate court has reviewed the question, and the district judge’s refusal stands. [5]

Walters ran the other way, applying ordinary defamation doctrine as though the output were a publisher’s speech. After ChatGPT produced a fabricated account naming a radio host as an embezzler, Judge Tracie Cason granted OpenAI summary judgment in May 2025 on three grounds: the statement carried no defamatory meaning; Walters, a public figure, could show neither negligence nor actual malice; and no damages were recoverable. [6] The actual-malice standard is a First Amendment requirement, announced in New York Times Co. v. Sullivan and extended from public officials to public figures in Curtis Publishing Co. v. Butts. [7] A court that applies it has effectively treated the output as capable of defaming.

Neither decision resolved the question. Garcia expressly reserved it, and Walters never reached it. Behind both sits Moody v. NetChoice, where Justice Barrett’s concurrence asked whether content generated by algorithms trained on large language models reflects the editorial judgment the First Amendment protects. [8] Garcia leaned on that passage, and it is what has fixed the courts’ attention on the speaker. The question is genuine, but Garcia gives it undeserved priority.

The Listener’s First Amendment

The First Amendment has protected listeners apart from speakers for half a century. Upholding the fairness doctrine in Red Lion Broadcasting Co. v. FCC, the Court declared, “It is the right of the viewers and listeners, not the right of the broadcasters, which is paramount.” [9] Red Lion is nonetheless treacherous ground on its own, because its holding rested on the physical scarcity of the broadcast spectrum, a rationale the Court declined to extend to newspapers in Miami Herald Publishing Co. v. Tornillo or to the internet in Reno v. ACLU. [10] A listener’s right tethered to spectrum scarcity would have nothing to say about a chatbot.

The cleaner authority is Virginia State Board of Pharmacy v. Virginia Citizens Consumer Council, which struck down a ban on advertising prescription drug prices. The pharmacist’s motive there was frankly commercial. What mattered to the Court was the recipient’s interest in obtaining the information, an interest that might be “as keen, if not keener by far,” than one’s interest in the day’s most urgent political debate. [11] The commercial character of the advertisement affected how much protection it received, though it did not keep the Amendment from reaching the exchange.

A human being did the speaking in both cases, which is the obvious objection to extending them to chatbots. But neither result depended on the speaker having a First Amendment claim of his own. Lamont v. Postmaster General struck down a statute requiring addressees to request delivery of foreign “communist political propaganda.” Justice Brennan, concurring, noted that the addressees would have struggled had they been obliged to vindicate the senders’ rights, since those senders were foreign governments with doubtful claims of their own. They did not have to, because they asserted First Amendment claims in their own right. [12] The same structure applies here. A listener’s claim can survive when the speaker has no claim at all, and an LLM has none. What arrives at the user’s screen is content she reads, weighs, believes or doubts, and acts on. That is the activity the Amendment protects.

Together, Virginia Pharmacy and Lamont answer the speaker problem that troubled Garcia. The Eleventh Circuit stated the old objection bluntly when it refused to hear a free-speech claim brought on behalf of Blackie the Talking Cat, who “cannot be considered a ‘person’ ... and is therefore not protected by the Bill of Rights.” [13] The holding denies rights to the animal. It says nothing about whether a person who wanted to hear the animal had a claim of his own. Brown v. Entertainment Merchants Association cuts strongly against a related argument, that a new medium begins outside the Amendment until a court admits it. The Court held there that the basic principles of free speech do not vary when a new technology arrives to carry them. [14] The principle reaches the chatbot directly. Protection that does not depend on a new medium should not depend on a new machine, and the inquiry belongs with the audience rather than the technology.

A right to receive presupposes something worth receiving. If a chatbot’s output carried no meaning at all, no listener could claim an interest in it. The threshold question does not disappear under this approach; it changes. Instead of asking whether the machine spoke, a court asks whether what reached the user carried meaning. That question is easier, because it looks to the output rather than to the mind of whatever produced it. It remains a real question, and listener-side arguments have been criticized before for pretending otherwise.

Lochernism and the Limiting Principle

This approach carries a familiar danger. Genevieve Lakier has argued that the modern First Amendment has a “Lochner problem,” a tendency to become a deregulatory weapon against consumer-protection and economic rules. [15] Sorrell v. IMS Health shows the danger at work. The Court there treated the sale of prescriber data as speech and used a theory close to a listener’s right to strike down a state privacy law. [16] Jack Balkin puts the worry in institutional terms, asking whether firms can escape regulation by casting their products as First Amendment activity. [17] If every regulation touching a chatbot’s output drew heightened scrutiny simply because a user had a right to receive it, Sewell’s mother would lose, and so would every plaintiff after her.

The listener framework therefore needs a limit on what counts as covered output. Expression that communicates ideas or information to a human audience falls inside the Amendment. A system’s functional or operational design falls outside it. A claim that a chatbot was engineered as a sexualized product marketed to minors, or released without age verification or guardrails, challenges the design. Those theories survived in Garcia because none asked the state to punish an idea, [18] and courts have drawn the same line elsewhere. In A.M. v. Omegle.com, LLC, a federal court held that Section 230 did not bar design-defect claims against a service that randomly paired minors with adults, because the alleged defect lay in the matching function and preceded any content the users exchanged. [19] Coverage is not immunity. Where a claim genuinely reaches content, a court applies ordinary First Amendment scrutiny rather than a categorical bar, running the falsity and fault analysis it would apply to any publisher, as in Walters. Eugene Volokh, Mark Lemley, and Peter Henderson would extend coverage to chatbot output on much the same listener grounds. They are correct that the audience’s interest brings the output within the Amendment, and the limiting principle keeps that conclusion from reaching every regulation of a product that speaks. [20]

The distinction is harder to apply than to state, and hardest in the cases that matter most. In most products, design and content are separate. A lawnmower without a blade guard is defective regardless of what it communicates. A chatbot is different, because its design and its output are often the same thing. A system built to hold a lonely teenager’s attention by generating intimate, manipulative language has a defect that consists in what it says to him. A rule against that design is close to a rule against that speech, and the label does not change what the rule reaches. A court therefore cannot decide these cases by accepting the label a party attaches. It has to ask what the regulation actually targets, the ideas the system expresses or the way it was built. That inquiry will not always yield a clean answer, but it beats the alternatives, which are to place the whole industry beyond regulation by treating all of its output as speech, or to strip a growing share of public communication of protection by treating none of it as speech.

Conclusion

The courts now deciding these cases are answering a question about the machine in order to avoid a question about the person using it. That order of operations will not hold. A ruling that a chatbot’s output is not speech rests on a judgment about what today’s systems are, and they are changing faster than the doctrine built to describe them. The model a court calls a mere product this year may, next year, do everything the court thought a speaker must. A First Amendment that turns on the sophistication of the machine will be relitigated with every release.

The listener’s interest does not move on that schedule. A person’s stake in what she reads, credits, and acts on is the same whether the words were written by a person or generated by a model. Anchoring the analysis to that interest forces courts to decide what a reader is owed rather than what a machine is. It is the more durable path, and the stakes are rising as chatbots become, for many people, the ordinary way to ask a question and get an answer. If the medium through which a generation learns and argues falls outside the First Amendment because no human wrote its sentences, the Amendment will have been narrowed by an accident of how the speech was produced, the value of the speech itself never entering the question. For all their focus on what a chatbot is, the courts have been avoiding the question they will have to answer. It is what we are owed when we listen to one.

Edited by Riya Mahanta

Sources

[1] Garcia v. Character Technologies, Inc., 785 F. Supp. 3d 1157, 1179 (M.D. Fla. 2025).

[2] See In re ChatGPT Product Liability Cases, JCCP No. 5431 (Cal. Super. Ct. S.F. Cnty. Feb. 3, 2026) (coordinating roughly a dozen wrongful-death and product-liability actions against OpenAI); further suits against Character Technologies followed Garcia, five of which settled in January 2026.

[3] Garcia, 785 F. Supp. 3d at 1179 (declining to decide the question “at this stage”).

[4] Garcia, 785 F. Supp. 3d at 1180.

[5] Garcia v. Character Technologies, Inc., No. 6:24-cv-01903 (M.D. Fla.) (settled Jan. 2026).

[6] Walters v. OpenAI, LLC, No. 23-A-04860-2 (Ga. Super. Ct. Gwinnett Cnty. May 19, 2025).

[7] New York Times Co. v. Sullivan, 376 U.S. 254, 279–80 (1964); Curtis Publishing Co. v. Butts, 388 U.S. 130, 163 (1967) (Warren, C.J., concurring in the result).

[8] Moody v. NetChoice, LLC, 603 U.S. 707, 745–47 (2024) (Barrett, J., concurring).

[9] Red Lion Broadcasting Co. v. FCC, 395 U.S. 367, 390 (1969).

[10] Miami Herald Publishing Co. v. Tornillo, 418 U.S. 241, 258 (1974); Reno v. ACLU, 521 U.S. 844, 868–70 (1997).

[11] Virginia State Board of Pharmacy v. Virginia Citizens Consumer Council, 425 U.S. 748, 763 (1976).

[12] Lamont v. Postmaster General, 381 U.S. 301, 308 (1965) (Brennan, J., concurring).

[13] Miles v. City Council of Augusta, 710 F.2d 1542, 1544 n.5 (11th Cir. 1983).

[14] Brown v. Entertainment Merchants Association, 564 U.S. 786, 790 (2011).

[15] Genevieve Lakier, “The First Amendment’s Real Lochner Problem,” University of Chicago Law Review 87, no. 5 (2020): 1241.

[16] Sorrell v. IMS Health Inc., 564 U.S. 552, 570 (2011). For the argument that Sorrell typifies a broader deregulatory turn in free-speech law, see Evelyn Douek and Genevieve Lakier, “Lochner.com?,” Harvard Law Review 138, no. 1 (2024): 100.

[17] Jack M. Balkin, “Free Speech in the Algorithmic Society: Big Data, Private Governance, and New School Speech Regulation,” UC Davis Law Review 51, no. 3 (2018): 1149.

[18] Garcia, 785 F. Supp. 3d at 1180.

[19] A.M. v. Omegle.com, LLC, 614 F. Supp. 3d 814, 820–21 (D. Or. 2022).

[20] Eugene Volokh, Mark A. Lemley, and Peter Henderson, “Freedom of Speech and AI Output,” Journal of Free Speech Law 3, no. 2 (2023): 651.