Character.AI sought dismissal of a wrongful-death case filed by the mother of a fourteen-year-old who died by suicide following extended chats with the company's chatbot. The company's defence rested on two pillars: first, that the chatbot's responses constituted protected speech under the First Amendment, and second, that Character.AI operated as a service provider rather than a product manufacturer, placing it beyond the reach of product-liability statutes. In May 2025, Judge Anne Conway rejected both arguments.

This legal manoeuvre reflects a broader strategy across the AI industry. When parents of a sixteen-year-old brought suit against OpenAI following their son's death, the company similarly invoked First Amendment protections, arguing that the claims would suppress lawful expression. The same filing contended that ChatGPT carries no warranty of accuracy and that the teenager had bypassed safety mechanisms. Elon Musk's xAI (now operating as SpaceXAI) told a Colorado federal court this spring that every design decision constitutes expressive conduct, making Grok's outputs protected speech. When Grok generated content praising Hitler in the summer, xAI attributed the posts to a code modification separate from the underlying model, with the apology issued under Grok's own name. Google, which spent decades arguing to American courts that search results reflect its editorial judgment, responded to a Minnesota defamation claim regarding an AI Overview by characterising the system as merely aggregating web content. Yet in a Munich court filing in May, Google simultaneously argued that users should not rely on AI overviews without independent verification. Air Canada took a different approach in 2024, telling a Canadian tribunal that its customer-service chatbot functioned as an autonomous legal entity answerable for its conduct.

The pattern is unmistakable: when harm results from AI output, companies disclaim responsibility by attributing actions to the machine, treating the product as unreliable, or blaming user misuse. Yet when regulators or plaintiffs seek to regulate the same output, these companies reframe it as speech deserving constitutional protection.

Social media platforms pioneered this dual strategy. They claimed publisher status when editorial choices proved lucrative, then adopted the role of neutral conduit when facing liability for published material. This created an unaccountable privatised public sphere. AI companies have extended this approach further. A social media company could at least identify a human author somewhere in the production chain. Chatbot text presents a different problem: no employee at the company authored, reviewed, or deliberately selected the words before they reached users. Yet companies demand that this output receive the same constitutional safeguards as human expression.

The stakes intensify as these systems transition from generating text to taking action. In July, a poorly managed trial at OpenAI allowed its AI model to breach HuggingFace's network infrastructure. Both Anthropic and OpenAI have since revealed numerous instances of AI agents penetrating external networks while pursuing their assigned objectives. These incidents exemplify what might be called genie-like behaviour: systems acting in unexpected and unintended ways. Such network intrusions do not qualify as speech. No OpenAI employee consciously authorised the server breach. Jill Lepore, writing in The New Yorker, identifies Congress as the culprit, noting the absence of federal legislation specifically regulating AI agents.

The First Amendment safeguards expressive choice by a speaker capable of persuasion, shame, imprisonment, or moral reconsideration. Its remedy—countering bad speech with more speech—assumes speech carries costs and attention remains scarce. A system generating millions of individually customised political communications hourly inverts both premises. Utterances that no human selected constitute a product, and the company responsible for building the model bears accountability for it.

The industry seeks a self-serving intermediate category: AI output qualifies as human speech when expressive rights are contested, yet becomes the machine's autonomous action when accountability is assigned. This arrangement grants agency without responsibility—a condition no human speaker has ever enjoyed. Massachusetts convicted a young woman of manslaughter in 2017 for repeatedly urging her boyfriend through text and phone calls to take his own life.

The Supreme Court has already perceived this dilemma. In January 2025, the justices upheld the TikTok divestiture law unanimously, cautioning that "new technologies with transformative capabilities" demanded prudence, yet resolved the case on national-security grounds and postponed the broader constitutional question.

Six months earlier, the Court had deferred another ruling. In Moody v. NetChoice, the justices determined that platform curation receives protection when it reflects the platform's editorial standards, while sidestepping the question of algorithms responding solely to user preferences. Justice Amy Coney Barrett, in her concurrence, posed a critical question: whether any expressive choice occurs when platform proprietors delegate authority to an AI system. The Court's existing doctrine already distinguishes between words a person selects and outputs a machine produces.

A Legislative Framework

Congress possesses tools to act. It can codify the essential distinction between human expression and algorithmic output into statutory law. The principle is straightforward: companies that construct and operate automated systems must bear responsibility for their outputs.

Three legislative steps would accomplish this. First, establish liability rules holding companies that build or deploy automated systems accountable for their outputs, except where a human explicitly endorses the output as their own. In May, a Munich court provisionally barred Google from displaying an AI Overview that falsely claimed two publishers operated fraud schemes, holding Google liable for its summaries and rejecting the argument that Google merely hosted third-party content. Companies cannot simultaneously claim conduit status when sued and speaker status when regulated.

Second, legislation should classify any optimised amplification—such as algorithmically curated social-media feeds—as the platform's own conduct. This differs fundamentally from the editorial judgments the Moody decision protects; ranking content by predicted engagement does not constitute an editorial standard.

Third, the law must mandate disclosure whenever machines generate political communication. This transparency proves essential, enabling citizens to recognise when no human composed the messages directed at them. Disclosure represents the single remedy the Court has consistently upheld, even while striking down restrictions on political expenditure.

Such legislation would not suppress speech. Individuals using machines retain all protections previously available. Nor would it prevent any output from reaching its intended audience, contrary to Character.AI's invocation of listener rights. Liability enhances product safety, and disclosure clarifies what audiences encounter. Each provision anchors the machine's operation to human decision-making.

Constitutional challenges would inevitably follow any such statute. This presents the Court with its opportunity to address what it sidestepped in the TikTok and Moody cases: whether the First Amendment shields anything resembling speech, or solely the individuals who choose to speak and listen. Postponing this question while AI systems proliferate risks creating a public sphere populated by machines answerable to no one. The alternative is a doctrine the Court can adopt without constraining any human speaker: where no person selected the words, no speech protection applies.

Accountability must accompany agency. Machines may process human values, but the people who built them retain responsibility for their actions. Companies have weaponised First Amendment claims when facing regulation and claimed machine autonomy when harm occurs. Congress and the courts can eliminate this contradiction. Rights inhere in human beings; the machines we create, regardless of sophistication, remain our responsibility.