AI and Authorship: The Modern Battlefield

Introduction

Ask any lawyer who owns a photograph, and you’ll get an answer before you finish the question: the photographer, unless a contract says otherwise. Ask who owns a paragraph a chatbot just spat out, and watch the room stall. Is it the company that built and trained the model? The software itself, in some new legal sense nobody has quite worked out? Or the person who typed the prompt? None of the three obvious candidates is entirely satisfying, and yet each keeps surfacing as a plausible owner in the disputes now working their way through courts and copyright offices.

This piece spends a little time on background law, not because it’s the hard part, but because you need it before the real fight makes sense. The interesting question isn’t whether a machine can hold copyright — it can’t; courts settled that already. It’s what happens once that’s off the table: when a judge actually has to choose among the developer, the system, and the human typing prompts, what does the law say, and where does it go quiet?

The Settled Background: No Machine Is an Author

Start with what nobody argues about anymore. The Copyright Office’s January 2025 report drew a fairly hard line: work with no human creative input can’t be registered, and typing a prompt by itself doesn’t make you an author, because the same prompt can produce wildly different outputs from one run to the next.[1] Thaler v. Perlmutter pushed that principle to its limit — an applicant who disclaimed any human involvement whatsoever — and lost, all the way up through the Supreme Court’s denial of cert in March 2026.[2] The D.C. Circuit’s reasoning leaned on small but telling details: the word “author” is woven through the Copyright Act in ways that only make sense for a human — inheritance provisions, a copyright term tied to a human lifespan, signature requirements for a valid transfer.[3]

Courts had actually seen something like this before. In Naruto v. Slater, the Ninth Circuit held that a macaque who pressed a camera shutter couldn’t hold copyright in the resulting photo, because words like “children” and “grandchildren” in the statute only make sense applied to people.[4] Swap the monkey for a neural network and the logic barely needs adjusting.

The input side of the problem — what a model may lawfully train on — is messier. Bartz v. Anthropic called training on lawfully acquired books “exceedingly transformative” fair use, while treating the company’s separately retained library of pirated copies as its own, unprotected problem.[5] Kadrey v. Meta reached a similar conclusion through a different route, though it left the door open to a “market dilution” theory for some future plaintiff with better evidence that AI output is displacing demand for the very works it learned from.[6] Thomson Reuters v. ROSS Intelligence went the other way, on facts where the output competed directly with the copyrighted material rather than merely absorbing its style.[7] None of that, though, touches the real question here: once a work clears the originality bar, who gets to claim it?

The Real Fight: Developer, Machine, or Prompter?

Three parties show up in nearly every AI-authorship dispute now working through a court or a copyright office, and each has a genuinely different theory behind its claim.

The AI System Itself — A Claim Going Nowhere
Start with the weakest of the three, if only because courts have already disposed of it. Dr. Thaler’s strategy was, in effect, to make the machine the author and himself the owner by extension, through something like a work-for-hire theory — treating his “Creativity Machine” as an employee. The D.C. Circuit wasn’t having it: work-for-hire reassigns authorship that a human already holds; it doesn’t manufacture authorship where none existed.[8] Feist Publications explains why this claim was always going to fail — copyright protects a modicum of creative choice, and a statute built around choice and judgment doesn’t have an obvious slot for something that makes neither.[9] Whatever future AI comes to resemble, the statutes on the books simply weren’t written with a non-human claimant in mind, and no court has been willing to read one in.

The Developer — A Claim Resting on Contract, Not Copyright
The company that built and trained the model holds a much stronger practical position than a legal one. It owns the code, the weights, usually the infrastructure — all separately protectable, all distinct from whatever the model later generates for a user. But owning the tool isn’t the same as owning what someone else makes with it, any more than a camera manufacturer owns its customers’ photographs. Developers who want a claim on outputs generally get there by contract, not doctrine: most major providers’ terms of service actually assign ownership the other way, handing users whatever interest exists in their own generated output, subject to a license the provider keeps for itself.[10] That’s a business choice, not a legal requirement, and it could flip. If a provider instead tried to keep ownership for itself — treating a user’s prompt as raw material fed into its own creative process — courts would probably have to run something close to the Reid analysis.[11]

The Prompter — Where the Real Litigation Is
This is where things actually get contested. The Copyright Office’s position is that a prompt alone isn’t enough — you’re choosing words, not brushstrokes, and the system’s own randomness breaks the chain of control the statute wants. But the same report conceded that AI-assisted work isn’t automatically excluded. A person who exercises real creative control — revising heavily, arranging generated pieces into something larger, or iterating with enough specificity to effectively dictate the result — can still claim authorship over their own contribution. The Office’s own call on Zarya of the Dawn drew that line in practice: it registered the human-written text and the human arrangement of panels, while carving out the individual AI-generated images as unprotectable.

Jason Allen’s pending case pushes directly on that boundary. Having won a state art-fair prize with a Midjourney image, he argues in his summary-judgment motion that dozens of iterative prompt revisions amount to exactly the kind of direct creative involvement the Office says should count — not the single-shot prompting it was worried about when it wrote the guidance.[12] Nobody knows yet how a federal court will rule, and that uncertainty is the point: the prompter’s claim isn’t dead the way the machine’s is, but it isn’t automatic either, the way a photographer’s claim to their own photo would be. It has to be earned on the facts.

Lay the three claims next to each other and a pattern shows up. The machine’s claim fails outright, cleanly, probably for good absent an act of Congress. The developer’s claim mostly isn’t being pressed right now, because giving ownership away is commercially easier than fighting for it — though that calculation could flip if outputs become valuable enough on their own. The prompter’s claim is presently being litigated in Allen v. Perlmutter. Rather than being resolved through a rigid bright-line rule, questions of AI authorship are more likely to be assessed along a spectrum of human creative involvement, ranging from minimal prompting to sustained and meaningful creative direction. Ultimately, the determination of authorship is likely to depend on the particular facts of each case, especially the extent to which the human user exercised creative judgment and control over the expressive elements of the final work.

India, and the Cyber-Law Layer Sitting Alongside All of This

India’s statute complicates things in an interesting way, because it already assigns authorship of a “computer-generated” work to whoever “causes the work to be created” — language that reads, decades early, like it was written for exactly this dispute.[13] Whether that phrase means the prompter, the developer, or someone else hasn’t been tested in an Indian court. The pending case that might test it, ANI Media v. Open AI, is actually built around training-data infringement rather than output ownership, and it raises a separate jurisdictional question about whether Indian courts can even reach a model trained and hosted abroad.[14] The government’s own December 2025 working paper proposed a licensing framework for training data, not an ownership rule for outputs, so the developer-versus-prompter question stays open under Indian law, for now.[15]

A separate layer of regulation runs alongside all this and doesn’t depend on who owns the copyright at all. India’s February 2026 amendments require synthetic content to be clearly labelled regardless of who claims authorship, and they narrow the safe harbour available to platforms that let unlabelled content circulate.[16] The EU’s AI Act imposes its own transparency duties independent of the copyright analysis, and China pairs labelling requirements with direct provider liability.[17] A developer, a model, and a prompter could resolve their three-way ownership dispute completely and still owe separate disclosure obligations under this other track of law.

Conclusion

The machine’s claim to authorship is dead, and Thaler buried it about as thoroughly as one case can. What’s still open — and what’s actually driving litigation and policy work through 2026 — is the contest between developer and prompter, a contest being settled today mostly by contract, in providers’ terms of service, rather than by copyright doctrine that hasn’t caught up yet. The Copyright Office has sketched a rough outline of where the prompter’s claim starts to succeed — somewhere past a single prompt, on the far side of real creative control — and Jason Allen’s case, along with whatever the Delhi High Court eventually does with language already sitting in India’s 1957 Act, should start filling in the details. Until then, the practical advice for anyone building or using these tools is unglamorous: keep a record of exactly how much creative control went into the final output. In every jurisdiction covered here, that record — not the prompt, and not the model — is what a court or copyright office will eventually want to see.

References

[1] U.S. Copyright Office, Copyright and Artificial Intelligence — Part 2: Copyrightability, 18 (Jan. 2025).
[2] Thaler v. Perlmutter, 687 F. Supp. 3d 140 (D.D.C. 2023), aff’d, 130 F.4th 1039 (D.C. Cir. 2025), cert. denied, No. 25-449 (U.S. Mar. 2, 2026).
[3] Thaler v. Perlmutter, 130 F.4th 1039, 1041, 1045–46 (D.C. Cir. 2025).
[4] Naruto v. Slater, 888 F.3d 418 (9th Cir. 2018) (holding that the animal lacked statutory standing to sue under the Copyright Act).
[5] Bartz v. Anthropic PBC, No. C 24-05417 WHA, 2025 WL 1741691 (N.D. Cal. June 23, 2025).
[6] Kadrey v. Meta Platforms, Inc., No. 23-CV-03417-VC, 2025 WL 1752484 (N.D. Cal. June 25, 2025).
[7] Thomson Reuters Enter. Ctr. GmbH v. ROSS Intelligence Inc., 765 F. Supp. 3d 382 (D. Del. 2025).
[8] Thaler, 130 F.4th at 1049–51.
[9] Feist Publ’ns, Inc. v. Rural Tel. Serv. Co., 499 U.S. 340, 345 (1991) (requiring only a modicum of creativity for originality, but requiring it nonetheless).
[10] See, e.g., OpenAI, Terms of Use § 3(a), https://openai.com/policies/terms-of-use (effective Jan. 1, 2026); Midjourney, Terms of Service § 4, https://docs.midjourney.com/docs/terms-of-service (effective May 27, 2026) (assigning to users, subject to stated conditions, any ownership interest the provider may have in generated output).
[11] Cmty. for Creative Non-Violence v. Reid, 490 U.S. 730, 751–52 (1989) (setting forth the common-law agency factors for determining whether a hired party is an employee for purposes of the work-made-for-hire doctrine).
[12] Allen v. Perlmutter, No. 1:24-cv-02665-WJM (D. Colo. filed Aug. 25, 2025) (pending; plaintiff’s motion for summary judgment filed Aug. 25, 2025).
[13] The Copyright Act, No. 14 of 1957, §§ 2(d)(vi), 52 (India).
[14] ANI Media Pvt. Ltd. v. Open AI Inc., CS(COMM) 1028/2024 (Delhi H.C., India) (pending).
[15] Dep’t for Promotion of Indus. & Internal Trade, Working Paper on Generative AI and Copyright: Part I (Dec. 8, 2025).
[16] Information Technology (Intermediary Guidelines and Digital Media Ethics Code) Amendment Rules, 2026, G.S.R. No. 120(E), Gazette of India, Extraordinary, pt. II, sec. 3 (Feb. 20, 2026).
[17] Regulation (EU) 2024/1689 of the European Parliament and of the Council of 13 June 2024 laying down harmonised rules on artificial intelligence, 2024 O.J. (L 1689).