An abandoned railway track surrounded by lush greenery ends in a dark, cavernous tunnel. This AI-generated image, titled “A Recent Entrance to Paradise,” gained fame in 2018 when Steven Thaler attempted to register it with the U.S. Copyright Office. Thaler listed his AI system, the Creativity Machine, as the work’s sole “autonomous” author. The Copyright Office rejected his application, and Thaler appealed in federal court and lost. The U.S. Court of Appeals for the District of Columbia Circuit held that authorship under the Copyright Act belongs exclusively to human beings, and that no machine, however sophisticated, qualifies as an “author” in any constitutional or statutory sense.

Now imagine Thaler turns his Creativity Machine to a different goal: electioneering. He abstractly instructs the Machine to autonomously generate, post, and A/B-test thousands of political attack messages across social platforms, optimizing for viral engagement. Thaler only sets a high-level goal to “take down all politicians who want to regulate AI” and the agentic system does the rest: including deciding who to target, what words, images, and videos to post, and which publications to select. A state legislature, alarmed at the scale of what’s coming, bans autonomous AI electioneering. Thaler sues again, this time claiming his Creativity Machine’s outputs are protected speech.

The two cases involve different legal doctrines. But they share a threshold question: When should the law protect machine-generated outputs that lack any meaningful human expressive contribution? Copyright law has spent nearly a decade working out an answer. First Amendment law has not. Could the Copyright Office’s approach to machine authorship offer a model for First Amendment doctrine, one that reins in agentic AI outputs and slop farms without chilling legitimate AI-assisted expression for humans?

Historically, the legal category of speech has never been stable. As Jennifer Petersen has documented, it has been actively remade with every new communication technology, from silent film to radio to computer code. Across those shifts, speech remained nominally tethered to human agency, but in an increasingly attenuated sense. Petersen calls this a move toward a “posthuman conception of speech,” where messages rather than persons become the locus of legal protection. That approach has already expanded corporate speech rights in ways many find troubling. Generative AI pushes the logic even further and leaves First Amendment doctrine without a clear account of who, if anyone, is speaking.

The First Amendment has generally dealt with human beings generating acts of protected expression. Landmark cases from the Vietnam-era jacket in Cohen v. California to the civil rights defamation case in New York Times Co. v. Sullivan to the Klan’s cross-burning in Brandenburg v. Ohio, all centered on concerns about human choices to say specific things. Even the corporate speech and right-to-listen cases, from Boston v. Bellotti to Citizens United v. the FEC and Martin v. City of Struthers through Virginia State Board of Pharmacy v. Virginia Citizens Consumer Council to Sorrell v. IMS Health Inc., involved corporations as vehicles for human expressive decisions, with human audiences who were interested in listening to human speakers.

That premise is now under pressure. Generative AI produces content at scale that is increasingly disarticulated from human involvement. Agentic systems coordinate tasks across platforms with minimal oversight. AI slop, or mass-produced synthetic content, is now the majority of new material posted to the open web. If legislatures try to regulate this in the typical domains of electioneering, consumer protection, and defamation, they will likely face a wave of First Amendment challenges from AI developers and operators claiming their machines’ outputs are protected speech. Call it slopigation: endlessly scaling litigation over AI content that no human directly expressed.

Without a threshold test asking whether any inherently human expressive contribution is present, courts have no principled way to sort protected speech from autonomous machine outputs. Courts would be mired in epistemologically complex and administratively unmanageable questions about whether agentic AI systems can be said to have viewpoints, and if so, whose they are. Worse, it could threaten to eliminate many forms of AI regulation.

So what could be learned from the copyright approach? When confronted with a flood of AI-generated material, the Copyright Office developed a two-part threshold test. The first factor evaluates the amount and type of human creative contribution. Here, the Copyright Office drew from a famous case of an 1882 photo shoot of Oscar Wilde, where Napoleon Sarony posed Wilde on a couch, dressed in dark velvet and silk stockings. A lithographic company later printed the image on trade cards without permission, claiming there was no copyright protection because the “author” of the image was a machine—the camera—not the human taking the photograph. The Supreme Court disagreed. Sarony had posed his subject “so as to present graceful outlines” and selected “costume, draperies, and other accessories.” A human who shapes a work as “the production of his own genius or intellect” is an author, while one who merely describes to a system what the work should do or look like is not.

The second factor focuses on output predictability. The Copyright Office distinguishes tools that extend a human’s creative choices from systems that generate outputs the human could not have predicted from her input. A Photoshop paintbrush extends a human’s choice predictably. Generative AI produces outputs the human couldn’t have anticipated from her prompt. Predictability matters because it tracks the locus of expressive contribution: The more the output predictably flows from human choices, the more the human is the author. If the output is the system’s own unpredictable recombination, there’s no human authorship.

How would these factors translate to a First Amendment context? Consider these four scenarios:

  • A journalist drafts an op-ed and then uses AI to tighten prose, suggest headlines, and check grammar. She makes express creative contributions and uses the tool predictably. This passes the speaker threshold.
  • An artist iterates across hundreds of prompts with unpredictable generative outputs, chooses between them and arranges them into a final composition. Express contribution flows through selection and arrangement rather than initial generation. In this case, only the selection and arrangement pass the speaker threshold.
  • A campaign operative deploys a swarm of autonomous agents to generate, post, and optimize thousands of message variants but makes no express contribution to any specific output, and the outputs are unpredictable from her high-level objective. This fails the speaker threshold.
  • A “pink-slime” operation publishes hundreds of AI-generated articles per day with nominal human review. Given the lack of express human contribution, there would be a strong presumption against passing the “speaker” threshold unless the operation can show its reviewers made meaningful editorial choices, such as through editorial selections, arrangements, or other substantive interventions.

This proposed framework doesn’t pretend to resolve the metaphysical question of whether AI outputs “really are” speech. It’s a more modest and tractable inquiry: Where is the human expressive contribution in the machine output, and how predictable was the output given that contribution?

One concern with this approach is that it might chill forms of AI-assisted human expression, particularly for those who lack the resources to litigate. In that case, the framework could include a presumption in favor of speaking. Some cases would be easy to satisfy, such as the journalist using AI to tighten prose. For more borderline cases, there could be a safe harbor. For example, once a speaker shows meaningful human contribution through chatlogs, prompt iterations, draft revisions, or other ordinary byproducts of creative work, a presumption of speaking attaches. Then the burden shifts to the opposing party to show that the human role was too attenuated or pretextual.

The presumption runs the other way for outputs produced at a pace or scale no human could meaningfully oversee, or when it flows entirely from automated, high-level commands. A platform that has trained an automated moderation system on detailed editorial criteria, audits its outputs, and can demonstrate that the system implements predictable expressive choices may rebut the presumption against speaking. This tracks Justice Barrett’s hypothetical in Moody v. NetChoice: AI processes that “implement human beings’ inherently expressive choice” remain protected, while AI processes given vague directives and allowed to run on their own do not.

There are three significant complications with the framework I’m suggesting. First, if corporations can invoke the First Amendment, why can’t AI systems? Citizens United and its predecessors might seem to be the logical go-to here. But the corporate speech cases arguably rested on the implicit premise that corporations aggregate and channel the expressive interests of identifiable human principals: shareholders, directors, officers. One interpretation is that the First Amendment protects the human expressive contribution that flows through the corporation, not the corporate form itself. The corporate form is a vehicle for human decisions and expression, while a swarm of autonomous AI agents is not.

Courts have also consistently declined to extend constitutional protection to non-people, including animals, in part because “unlike the human species, which has the capacity to accept social responsibilities and legal duties, nonhuman[s] cannot—neither individually nor collectively—be held legally accountable or required to fulfill obligations imposed by law.” This applies equally to machines, which cannot be held legally accountable in and of themselves: only the humans or corporations that own or administer them can. It’s consistent with copyright’s longstanding refusal to extend protection to non-human entities, including spirits, monkeys, and well-designed gardens.

This brings us back to Barrett’s concurrence in Moody v. NetChoice, where she questioned whether AI-based content moderation decisions would be First Amendment-protected. AI-based decisions could qualify, she suggested, if they “simply implement human beings’ inherently expressive choice ‘to exclude a message [they] did not like from’ their speech compilation,” for instance “to remove posts promoting a particular political candidate or advocating some position on a public-health issue.” But she contrasted this with a scenario in which “a platform’s owners hand the reins to an AI tool and ask it simply to remove ‘hateful content,’” and then she asks: “If the AI relies on large language models to determine what is ‘hateful’ and should be removed, has a human being with First Amendment rights made an inherently expressive ‘choice’?” Such autonomous AI-driven actions may be too removed from “human beings’ constitutionally protected right” to merit protection. This parallels the same concept of attenuation that the Copyright Office uses in its test for authorship.

At least one district court has applied Barrett’s approach in rejecting an AI company’s early-stage argument in favor of First Amendment rights for its fully automated chatbot. This also echoes the Copyright Office’s framework and copyright law’s broader “volitional conduct” doctrine, which acts as a threshold test for liability and denies it where infringing copies are made purely through automated machine activities.

Second, even without a human speaker, don’t listeners have a right to receive AI-generated content? As this argument goes, even if AI outputs lack a human speaker, human listeners may wish to receive them, and the First Amendment often protects the right to receive information. Cases such as Lamont v. Postmaster General, Virginia Pharmacy, Stanley v. Georgia, and Red Lion Broadcasting Co. v. FCC have often been read to protect access to expression regardless of the speaker’s interest. But the right-to-listen cases involved human speakers whose communications were being kept from human audiences. Where there’s no human speaker, it becomes more complex. You may enjoy the sound of ear-splitting construction noise or whale songs but that doesn’t necessarily implicate your constitutional rights. At least one court has specifically held that the right-to-listen does not apply to non-human speakers in the case of captive rhesus macaques monkeys, even if they are “willing speakers under the First Amendment” who regularly communicate “through vocalizations, facial expressions, head and limb movements” and other behaviors. AI outputs derived from training data are not, in the relevant sense, what other people are saying. They are statistical recombinations of past human expression to which listeners generally have direct access through other means.

Murthy v. Missouri closed off the broadest version of this argument. The Court confirmed that generalized listener rights claims unanchored to specific speakers and specific suppressed content don’t establish a concrete First Amendment injury. So a general right to listen to AI is unlikely to succeed. Specific humans would have to show their interest in specific AI content just to qualify for court standing to challenge AI regulations. The copyright parallel is instructive here too. The Constitution grounds copyright in promoting the “progress of science and useful arts,” a purpose that might seem to justify protecting machine authorship on listener-benefit grounds. Courts and the Copyright Office have rejected that move. If limiting authorship to humans is constitutionally sufficient to promote the progress of science, it is hard to see why limiting First Amendment protection to human speakers isn’t also constitutionally sufficient.

A third issue concerns data. For years, Sorrell has troubled regulators, information law scholars, and privacy advocates because its language appears to extend First Amendment protection broadly to information flows that eventually lead to recognized speech. Read maximally, Sorrell would protect almost any data eventually used to generate any protected output. In the context of agentic AI, this could mean almost any output or activity.

But Sorrell, like the corporate speech and listener rights cases, involved human-generated information: pharmacists’ records of doctors’ decisions, recorded and aggregated for human audiences. Extending that logic to machine-generated outputs would leave almost no room for any regulation of online content at all. A better reading, consistent with Barrett’s Moody concurrence, is that First Amendment protection tracks the presence of human expressive choices. It attaches where machine outputs predictably implement those choices and weakens as that connection attenuates.

Agentic AI risks redrawing the boundaries of speech in ways that favor machines over the individuals who might use them. A threshold anchoring First Amendment protection to human expressive contribution offers one doctrinal mechanism for resisting that tendency. It does not resolve what scrutiny applies once a human speaker is established, how courts should handle hybrid cases in which human and machine contributions are deeply entangled, or whether the line between human cognition and AI assistance will even be locatable in the next decade. Those are questions for another day.

But the courts need a threshold, and soon. The next wave of First Amendment litigation will not look like the last one. It will involve content that no human authored, accounts running on autopilot, and speakers who never spoke. Right now, there is no established approach to handle these cases. The Copyright Office had to develop a threshold test to deal with the influx of AI-generated content, and First Amendment doctrine can do the same. The Amendment protects human speech and association. It was not designed to shelter slop factories or swarms of autonomous agents from democratic regulation.