待翻譯：If AI Outputs Aren't Speech, Who Has to Prove They're Human?

AI 服務暫時不可用，以下為來源正文，待恢復後補全翻譯。

Becca Branum Meet The Authors Subscribe to Lawfare A growing body of legal scholarship argues that large language model (LLM) outputs are not “speech” under the First Amendment and therefore may be regulated without the scrutiny ordinarily applied to restrictions on expression. Garcia v. Character Technologies, a wrongful-death suit brought after 14-year-old Sewell Setzer III died following months of conversations with an AI character, resulted in one of the first judicial opinions addressing the constitutional status of chatbot outputs. Ruling on the company’s motion to dismiss, U.S. District Judge Anne Conway wrote that she was “not prepared to hold that [LLM] output is speech” and allowed product liability and negligence claims to proceed. Garcia settled in January 2026, but other suits continue with similarly disturbing facts alleging that chatbot outputs played a role in medical crises, violence, or providing sexualized content to minors. Following these suits, a bevy of federal and state proposals would change whether and how people can receive information from chatbots. Whether the no-speech position is doctrinally correct is one question. What it would take to administer that position is another. If human expression is protected but machine output is not, legal coverage may depend on whether a person created, selected, edited, or adopted—that is, knowingly put forward as one’s own—the words generated by the LLM. Yet ordinary text rarely reveals whether it was written by a person, generated by a model, or produced through some combination of the two. If human attribution determines whether the First Amendment applies, the law needs both an attribution standard and a default when attribution cannot be established. Administering that rule would mean sorting human from machine expression at scale—a task no existing tool can reliably perform. The likeliest substitutes, identity and personhood verification, establish who a speaker is or that a speaker exists—not who authored a given text. Either approach burdens the very expression the rule purports to leave protected. The burden of that uncertainty would likely fall more on users and readers than model developers. Enforcing a human-attribution rule across accounts, platforms, and disputed works requires some way to distinguish person from machine. Existing methods either cannot reliably distinguish between the two or demand more information about the speaker. The risks include reduced anonymity, greater surveillance, and pressure to use identity or personhood as a proxy for authorship. A consequentialist argument for excluding artificial intelligence (AI) output from First Amendment coverage should account for those costs when human involvement is mixed or disputed. The No-Speech Position The no-speech position argues that because no person stands behind a model’s words at the moment of generation, those words are not First Amendment-covered “speech.” The leading versions of the no-speech argument arrive at the same conclusion via different routes. Mackenzie Austin and Max Levy argue that machine output lacks “speech certainty:” No human speaker knows what a model will say as it says it. Peter Salib contends that AI output is the protected speech of no one—not the model, developer, or user. David Atkinson, Jena Hwang, and Jacob Morrison argue that frontier models lack communicative intent and therefore produce no speech at all. The appeal is understandable. If model outputs are treated as speech, laws regulating them may trigger heightened First Amendment review. That result can seem perverse for laws intended to address product safety, fraud, or discrimination. The impulse is not frivolous: It reflects a genuine worry that ordinary First Amendment doctrine might otherwise give constitutional shelter to large areas of automated system behavior. The approaches differ over when subsequent human interaction is enough to earn constitutional protection, but each makes human attribution the deciding factor. Benjamin Wittes framed the doctrinal pressure from the other direction. Where some no-speech scholars deny protection to machine output to avoid an unpalatable result, he argued that existing doctrine, followed faithfully, already extends protection to it. Writing in Lawfare, he arrived at what he called “the first machines with First Amendment rights.” Machines don’t have constitutional rights, but a company’s expressive rights are not the machine’s; they are the rights of people who own and direct it. But a court need not give rights to a machine to recognize that regulating its outputs may burden the rights of users, readers, or developers. The difficulty is not the theory but its administration. In an individual lawsuit, attribution may be manageable. A court can examine the prompts, editing, selection, product design, model operation, parties, and claims after discovery. A rule applied across the internet must make similar judgments without that record, often before anyone knows who contributed what. And the words themselves usually cannot resolve the question. The Distribution Problem The attribution question would not stay confined to lawsuits against developers. Most text reaches readers through intermediaries—such as platforms, hosts, search engines—and a no-speech rule changes their incentives too. Distributor liability for illegal or unprotected material can still raise constitutional concerns when it chills protected speech. But the most significant protection for online intermediaries is statutory: Section 230 of the Communications Decency Act, which bars treating a provider or user of an “interactive computer service” as the publisher of information provided by another “information content provider.” That shield does not depend on whether the underlying content is protected. Section 230 fits less comfortably when the service itself appears to generate the substance of a communication. An LLM generates new text in response to a prompt, blurring the line between hosting third-party content and creating one’s own. No court has decided whether chatbot output is the kind of “information provided by another” that the statute covers, and Section 230’s own authors have said it should not be. That shield could give way for AI-generated outputs from either direction. Courts may hold that Section 230 does not reach model output, or Congress may withdraw it. The pressure to do the latter is real and bipartisan: In December 2025, senators led by then-Sen. Lindsey Graham introduced the Sunset Section 230 Act following a similar House proposal the year before. A repeal would expose anyone who hosts online content to liability for it, but under a no-speech rule, that exposure would be highest for machine output. Every distributor would then have a concrete reason to know which items moving through its systems are machine-made and which users are human. Complete repeal aside, legislators have shown an interest in eliminating protection for AI under Section 230, which might also exclude AI-generated content from its shield. The risk does not vanish under a no-speech rule even if Section 230 survives. The statute limits liability for hosting another’s content; it does not limit what a state may demand of a platform’s own conduct. A legislature that placed unadopted machine output outside the First Amendment could regulate it directly—by barring platforms from carrying it, or by conditioning access on proof that a real person stands behind each post. None of this is settled law, but the incentive structure is clear: Make human attribution the line between protected and unprotected expression, and governments concerned about AI-generated disinformation, harassment, or foreign influence will push institutions to sort human from machine at scale. When Attribution Becomes the Test The attribution problem does not arise in every case. A law regulating a known chatbot provider’s conduct or requiring automated responses to identify themselves need not determine whether an unknown online speaker is human. The problem is harder when the source is disputed. Suppose a state requires machine-generated election advocacy to carry a conspicuous label, while exempting materially identical advocacy written or adopted by a person. Consider an automated agent that a campaign volunteer set running weeks earlier with a single general instruction—promote this candidate—and that now drafts and posts its own advocacy. A human set the process in motion with a communicative purpose, but no human wrote, selected, or adopted the specific post. If that distant instruction counts as adoption, the compelled label burdens human expression; if not, the no-speech theory makes the constitutional objection weaker. A classifier might flag model-like phrasing. Provenance might confirm an AI tool touched the file somewhere along the way. A personhood credential might confirm a human controls the account. What none of them can show is adoption—that a specific person stood behind this specific text. Enforcement then turns not just on detection, but on the legal default: whether ambiguous posts are treated as protected unless the government proves otherwise, or unprotected unless the speaker proves adoption. What the Tools Can Prove Institutions asked to make that sort have essentially three families of tools—text classifiers, content provenance, and behavioral analysis—plus a fourth, personhood credentials, discussed below. Each answers a real question. None answers the one a human-attribution rule poses. Text classifiers are the most direct option, but they are unreliable and answer the wrong question. OpenAI withdrew its own classifier in 2023 because of low accuracy. Researchers have also found that detectors disproportionately misclassify writing by non-native English speakers. Even a perfectly accurate classifier only tells you whether a text resembles model output—not how, if at all, a person was involved, as drafter, editor, or adopter. In a single enforcement action, that gap need not be decisive: An agency with subpoena and warrant power can compel the logs, account records, and testimony a classifier never sees. The difficulty is the routine, at-scale sorting the distribution problem invites, where a platform or rule-maker must separate human from machine expression across millions of items, often with only the words to go on. As mixed human-machine production becomes routine, a binary label becomes less informative. Provenance is better evidence of how a file was made and changed than of who’s responsible for what it says. The C2PA standard, developed by an industry coalition that includes Adobe and Microsoft, can attach a signed, tamper-evident history to digital content. By its own documentation, “Content Credentials,” the C2PA standard’s consumer-facing label that lets readers inspect a file’s recorded history, confirm that provenance data is tied to an asset and hasn’t been tampered with—it doesn’t vouch for whether the underlying claims are true. Behavioral systems, such as spam filters, bot detection, device fingerprinting, posting patterns, and network analysis, can identify likely automation. But they are probabilistic and adversarial, and they generally classify an account rather than particular words. The more confidence a system seeks, the more closely it must inspect the person or device behind the content. As Madeline Lamo and Ryan Calo observed in “Regulating Bot Speech,” requiring someone accused of operating a bot to prove otherwise m [truncated for AI cost control]