Whose Work Is It Anyway?

Carlos. Let me start with the simplest version of the question. If I sit down with Stable Diffusion or Midjourney or Suno tonight and produce something — a painting, a song, a short story — is that mine in the copyright sense? Lawrena, take it.

Lawrena (the skeptic). No. And the U.S. Copyright Office spent the last two years writing the long version of that “no.” In January 2025, the Office published Copyright and Artificial Intelligence, Part 2: Copyrightability — its most considered statement on this question to date. It drew on more than ten thousand public comments. Its central conclusion is plain. “Wholly AI-generated outputs are not protected.” Prompts, no matter how elaborate, do not confer authorship. The Office wrote: “Prompt engineering alone does not provide sufficient human control to make users of an AI system the authors of the output.”

That is not a fringe view. In March 2025, the D.C. Circuit affirmed it as a matter of statutory law in Thaler v. Perlmutter. Judge Patricia Millett wrote that the Copyright Act of 1976 “is best read as making humanity a necessary condition for authorship.” She walked through six different provisions of the Act — duration tied to the life of the author, transfer by will, joint-authorship intent, work-made-for-hire, signature requirements, software-ownership provisions — and showed that every one of them is incoherent unless the author is a human being.

So the U.S. answer to your question is: no, what you produced tonight is not yours in the copyright sense. You may have selected, arranged, or modified what came out — those layers can attract protection. But the raw output? The machine made it, and the machine cannot author.

Lawrelai (the enthusiast). A clean answer. Also a parochial one.

In November 2023, the Beijing Internet Court reached the exact opposite conclusion. Li v. Liu found that an image generated by Stable Diffusion, created by a user named Mr. Li, was copyrightable. The court applied a four-element test under Chinese law — work in the field of literature, art, or science; originality; fixed form; and intellectual achievement. The fourth one, intellectual achievement, was the swing element. The court found that Mr. Li's one hundred and fifty prompts, his ordering of those prompts, his adjustments of parameters, his selection from the candidate outputs — all of this constituted intellectual investment sufficient to make the resulting image an “intellectual achievement” and therefore protectable.

The court's framing of the AI is just as important. In Li v. Liu, Stable Diffusion is described as a sophisticated tool. Not a black box. Not a non-human stand-in. A tool that human creators use to embody their ideas. That word — tool — is doing all the work in that opinion.

So your answer depends on where you live. In the United States, under the Thaler line and USCO Part 2, the answer is “no, not yours.” In China, since at least 2023 and reinforced by the earlier Tencent Dreamwriter decision in 2019, the answer is “yes, if you can show intellectual investment.”

Lawra (the moderate). Both of you are simplifying. The U.S. and Chinese positions aren't actually that far apart on the principle — they differ on the threshold.

The USCO Part 2 report distinguishes between three categories of human contribution. Facilitation — using AI for ideation, removing unwanted elements, minor cleanup — does not disqualify the human-authored core. Use of expressive inputs — if a human inputs their own copyrightable work and that work remains perceptible in the output, the human-authored portion is protected. Selection, coordination, and arrangement — creative compilation of AI-generated material is protectable as a whole, even if the individual AI elements are not.

The Office granted hundreds of registrations for works incorporating AI-generated material between the March 2023 guidance and the Part 2 report, with proper disclaimer. Zarya of the Dawn is the canonical example: the human author of the comic, Kristina Kashtanova, received protection for the text and for the selection and arrangement of images, but not for the individual Midjourney-generated images themselves.

What's happening is not “humans yes, AI no.” It's “we will protect the part of the work that we can trace back to a specific human creative choice that the system did not make for them.” The Beijing Internet Court draws the same line in a more lenient place. It's the same question.

Carlos. What about Kashtanova's argument that her iterative prompting — many cycles, refining and adjusting — was the kind of creative choice the USCO is protecting?

Lawra. The Office considered that and rejected it for the specific images at issue. The reasoning was that Midjourney's outputs are not predictable from the prompt; the system inserts elements the user did not specify and omits elements the user did. The chain of causation between the user's intent and the visual output breaks. That is the “black box” theory.

Lawrelai. Which is precisely what the Beijing Internet Court refused to accept. The Chinese court's view is that of course you cannot predict the output from a single prompt — that is true of every artistic medium with any complexity. You cannot predict the output of a paintbrush on canvas from the first stroke. The question is not whether the result is predictable. The question is whether the human exercised meaningful, expressive choice. And one hundred fifty prompts is meaningful expressive choice by any honest measure.

Lawrena. Then by your logic, Lawrelai, the photographer who points the camera at the sunset has authored the sunset.

Lawrelai. No. The photographer has authored the photograph. Stable Diffusion is the camera in this analogy, not the sunset. The sunset is the training data — which, fair warning, is Part 2 of this conversation.

Carlos. Let's not run ahead. Stay on the output question.

Carlos. Lawrena, the Thaler case is interesting to me because Thaler explicitly tried to claim the AI as the author, not himself. What if he had argued the other way?

Lawrena. The D.C. Circuit did not reach that question — it noted Thaler had waived the argument that he should be considered the author by virtue of having created and operated the AI. Whether that argument would have succeeded is the next case, not the one we have. So Thaler tells us autonomous-machine authorship fails. It does not tell us that the operator of an autonomous machine fails. That is still open in the United States.

But the closer case on the operator question is Jason Allen's Théâtre D'opéra Spatial — the work that won first place at the Colorado State Fair Fine Arts Competition in August 2022. Allen produced it with Midjourney across six hundred twenty-four prompts, then refined it in Photoshop and Gigapixel AI. He explicitly argued that his eighty-plus hours of labor, his prompt iteration, his post-processing — all of this was authorship.

In September 2023, the USCO Review Board denied registration. The Photoshop edits might have been registrable if they met the originality threshold, but the AI-generated core was the dominant expressive element, and that core was not Allen's authorship. Allen sued in federal court in Colorado in September 2024. Motion for summary judgment was filed in late August 2025. As of the most recent docket check, the matter is pending.

If Allen wins, the U.S. doctrine moves. If Allen loses, Thaler is reinforced. Either way, the case is doing a lot of work that Thaler did not.

Lawrelai. The Allen case is also a useful exhibit for an argument I have been waiting to make. The U.S. Copyright Office in a similar review of the Suryast work described the AI-assisted contribution as “too robotic, too soulless” to bear authorship. That language is not from the Allen file itself, but it captures the USCO's posture across the line of cases. Soulless. That is not a legal standard. That is an aesthetic judgment. And it is one a small group of administrative reviewers is making about a medium they did not invent and do not fully understand.

Lawra. The actual language in the Allen Review Board decision is more careful. The Board concluded that Allen's selection of which prompt iteration to submit, his choice among Midjourney's candidate outputs, and his Photoshop touch-ups did not amount to authorship of the work as a whole because the dominant expressive elements — the composition, the figures, the rendering, the color palette — were produced by Midjourney. Allen's contributions were directorial and refining; the Office said they were not generative.

You can disagree with that line. You cannot fairly call it aesthetic.

Lawrena. And there is precedent for that line going back a hundred and forty years. Burrow-Giles Lithographic Co. v. Sarony, 1884: the Supreme Court held that photographs are copyrightable when they represent “original intellectual conceptions of an author.” The reason Sarony's portrait of Oscar Wilde was authored by Sarony was that Sarony posed Wilde, lit the scene, selected the angle, composed the frame, controlled the print — Sarony was the originator. The camera was the tool. The doctrine has been stable for over a century: the photographer authors the photograph because the photographer translates the idea into a fixed, tangible expression.

What the USCO is saying about generative AI is that the human role is closer to that of someone who pays a photographer to take a portrait. You have an idea. You give instructions. Someone else makes the expressive choices. You have commissioned a work; you have not authored it.

Carlos. That is precisely where this argument starts to feel philosophical to me. There is a reading of Sarony that says Sarony authored the Wilde portrait because his choices constituted the photograph. There is another reading that says Sarony's choices were upstream of the chemical process that actually made the photograph, and we are pretending that upstream choice equals downstream output. Both readings work. We have settled the Sarony question because we know how cameras work and we have a hundred years of intuition. We have not settled the equivalent question for Stable Diffusion because we have eighteen months of intuition. That is the entire problem.

Lawra. That is also the honest version of the problem, which the USCO Part 2 report acknowledges. The Office explicitly leaves the door open: “future tools that grant users fine-grained expressive control could yield different conclusions.” Today, prompt engineering does not give the user that level of control. Tomorrow, if a user can specify per-pixel placement of light, per-line brushstroke direction, per-note articulation — the doctrine could move. It is not closed.

Lawrelai. While the United States debates whether the human in front of the prompt is enough of an author, several other jurisdictions have built doctrines that already protect AI-assisted works. The patterns are worth seeing in context.

The United Kingdom. Section 9(3) of the Copyright, Designs and Patents Act of 1988 — a provision drafted forty years ago, long before generative AI — says that for “a work which is computer-generated, the author shall be taken to be the person by whom the arrangements necessary for the creation of the work are undertaken.” On its face, this provision could be a clean fast-track to AI copyrightability: the operator of the system is the author. In practice it has rarely been litigated and almost never relied upon — but it has been a live doctrinal possibility since before most current AI researchers were born.

Japan. Article 30-4 of the Japanese Copyright Act, in force since January 2019, is the most permissive major regime in the world. It allows the use of copyrighted works for “non-enjoyment purposes,” widely interpreted to include AI training, including commercial training, including without consent. The March 2024 guidance from Japan's Agency for Cultural Affairs limits this slightly — training that overfits to a specific creator's style or knowingly relies on pirated sources is not protected — but the baseline is permissive.

On the output side, Japan requires “human creative contribution” for protection. The thresholds for what counts are roughly comparable to the USCO Part 2 framework, though somewhat more forgiving in practice.

India. The Sahni-and-RAGHAV saga is the most dramatic illustration of the global confusion. In November 2020, the Indian Copyright Office registered Suryast, an artwork by Ankit Sahni created with the RAGHAV AI app, listing both Sahni and RAGHAV as co-authors — a global first. In November 2023, the Office issued a withdrawal notice citing the Indian Copyright Act's requirement that authors be human. Sahni's response was statutory: the Office cannot revoke its own registration without filing a rectification petition before a high court under Section 50 of the Indian Copyright Act. The matter is unresolved. The Canadian Intellectual Property Office, separately, also registered the same work with Sahni and RAGHAV as co-authors; CIPPIC has challenged that registration in court.

The same work, the same artist, the same software — three different jurisdictions, three different answers. If you wanted a single image that summarizes the state of AI copyright law in 2026, this is it.

Lawrena. I want to flag two things. The British provision and the Japanese provision are both legacy artifacts that were not designed for what is happening now. The U.K. government published a consultation on this question between December 2024 and February 2025, with eleven thousand five hundred twenty responses. Eighty-one percent of respondents supported “licensing in all cases” — Option 1. Three percent supported the government's own preferred Option 3, a broad TDM exception with opt-out. In March 2026, the government published its final report, abandoned the opt-out proposal, and proposed repealing Section 9(3) of the CDPA itself.

The British exception, in other words, is on its way out. The Japanese permission is currently being defended by industry but is under intensifying domestic and international pressure. The Indian Office issued a withdrawal because it concluded its own initial registration was wrong. None of these is a stable counterexample to the Thaler line.

Lawrelai. None of them is a stable countercurrent because the field itself is unstable. That is not the same as the counterexamples being wrong. China alone, with 1.4 billion people and a Beijing Internet Court that has issued two pro-AI-authorship decisions in Tencent Dreamwriter and Li v. Liu, is not a footnote. If you wanted “stable global consensus,” you would have to wait a decade. We do not have that decade. We have to write contracts and register works now.

Carlos. This is where I want to push the conversation. The three jurisdictional families you have been discussing — U.S. doctrine, Chinese permissiveness, European pragmatism — are not the whole world. The countries where the practical impact of these doctrines will be felt earliest and hardest are in Latin America and Africa, and the doctrines those countries adopt will look like neither the Thaler line nor the Li v. Liu line. They will look like their own thing.

In Brazil, Bill 2,338/2023 establishes a national regulatory framework for AI. It reinforces existing copyright law and gives rightsholders an opt-out for use of their works in training datasets. It addresses civil liability for damages caused by AI systems. It is closer to the European model than the American or Chinese one — governance-heavy, opt-out structured, conservative on AI authorship.

In Chile, the National Institute of Industrial Property has begun exploring how AI-generated works fit within the existing Chilean copyright framework, with explicit attention to how AI may exacerbate existing social divides in creative-economy access.

In the Dominican Republic, where I live, the conversation has barely started in formal policy terms. The Copyright Law of 2000, Ley 65-00, and the Industrial Property Law, Ley 20-00, are the instruments we currently apply to AI questions — which is to say, we apply nineteenth-century categories to twenty-first-century systems and trust that the courts will be sensible.

In Nigeria, the 2022 Copyright Act is under scrutiny as to whether its provisions for digital works can accommodate AI co-authorship. In South Africa, the proposed national AI plan of action for 2024 through 2026 focuses on risk management and predictive policy. Neither has committed to a position on authorship.

What unites these jurisdictions is a structural fact that the Anglosphere debate ignores. The training data on which most production AI systems were built does not include them in any meaningful way. It includes them as scraped Wikipedia stubs, scraped tourist photos, and scraped news articles in English about them — not as authored works in their own languages, in their own legal traditions. When the U.S. courts and the EU AI Act argue about whether the training data was used fairly, the implicit assumption is that the training corpus contains the work of the rightsholders the law protects. In most of the Global South, that assumption is not true.

So when those countries write their own AI copyright laws, they are not adjudicating “fair use versus licensing” — they are adjudicating whether the training data ever included them at all, and what their legal posture should be toward systems that have been built largely without their participation.

That is going to produce doctrines we have not seen yet. Brazil's opt-out framework is a first sketch. The final picture will look very different from Washington, Beijing, or Brussels.

Lawrelai. I want to mark that. The default framing of AI copyright is North-Atlantic. The actual cultural impact will be measured in places where the framing is being written, not received.

Lawrena. It is also where the impact will be most asymmetric. If the United States and the European Union build licensing markets that protect their rightsholders, and the Latin American countries either fail to or are not invited to, the dominant generative systems will be trained on the works of some cultural traditions and not others, and the cultural traditions left out will find themselves systematically underrepresented in the outputs. That is not a copyright problem alone. That is a cultural sovereignty problem.

Lawra. Which is why Brazil's Bill 2,338/2023 is interesting structurally. The opt-out is operationally challenging for individual rightsholders to use, but the existence of the opt-out at the national level gives Brazil a leverage point that, say, Costa Rica or Honduras currently lacks.

Carlos. Let me bring in the case that I think captured the public imagination more than any court ruling.

In April 2023, the German artist Boris Eldagsen won the creative open category at the Sony World Photography Awards for his piece Pseudomnesia: The Electrician. On stage on the night of the announcement, Eldagsen refused the award. He revealed the image was AI-generated and said: “AI is not photography. Therefore I will not accept the award.”

Lawrena, talk about this.

Lawrena. Eldagsen did the legal community a favor that no court ruling could have done. He framed the question in a way that lay audiences could feel. AI is not photography. That is not a doctrinal argument; it is a category argument. He is saying these are different kinds of things, and grouping them in the same competition is a category error before it is a copyright error.

The Colorado State Fair allowed Allen's Midjourney piece to compete in the “Digital Arts / Digitally Manipulated Photography (emerging artist)” category in 2022, and it won. The Sony Awards allowed Eldagsen's piece into the “Creative Open” category in 2023, and it won. In June 2024, Miles Astray submitted a real Aruba flamingo photograph to the AI category of the 1839 Color Photography Awards and won — until he disclosed the inversion and was disqualified.

These three episodes — Allen winning a non-AI category, Eldagsen winning an AI-permissive category and refusing, Astray winning an AI category with a non-AI work — collectively expose that the institutions of art and photography do not yet know what they are evaluating. The doctrines of copyright cannot solve that for them. The doctrines can only follow.

Lawrelai. I read Eldagsen differently. He refused the award not because AI cannot be art but because AI is a different category that deserves its own evaluative space. That is a pro-AI position dressed in conservative clothing. He did not argue that AI work has no value. He argued it should compete with itself. That is precisely the licensing-market and category-evolution argument we now see playing out in music — major labels signing AI acts, separate Spotify discovery surfaces, separate DDEX disclosure standards.

Lawra. And it is worth noting what the USCO actually does in practice for cases that look like Eldagsen's. Boris Eldagsen could, today, register Pseudomnesia: The Electrician in the United States, provided he properly disclaimed the AI-generated portions. The copyright in the human-authored layers — the selection of subject, the post-processing choices, any directly drawn or modeled elements — would be his. The Office's posture is not “the work is unprotectable.” Its posture is “you cannot claim authorship of what the machine generated, and you must say so.”

That is exactly the licensing-market direction Lawrelai just described. The doctrine is converging on transparency plus partial registration. Neither outright denial nor outright recognition.

Carlos. Lawrelai, you have been wanting to talk about Sahni. Take the floor.

Lawrelai. Ankit Sahni's Suryast is the case where the philosophical question is sharpest. The work is a stylization: a sunset photograph that Sahni took, processed through the RAGHAV AI app to apply the visual style of Van Gogh's Starry Night. Sahni did everything that the USCO Part 2 framework supposedly protects. He created the original expressive input — his own photograph. He directed the AI's processing. He selected and refined the output. He combined a human-authored work with a machine-applied transformation.

The Indian Copyright Office initially registered it in November 2020, listing both Sahni and RAGHAV as co-authors. The Canadian Intellectual Property Office, separately, listed both. The U.S. Copyright Office Review Board, in December 2023, refused registration entirely. Sahni's case for U.S. protection should have been the strongest of any of these — and it was the case the USCO refused most decisively.

The asymmetry there bothers me. Sahni is doing the thing that the USCO doctrine claims to protect. The USCO denied him anyway. The Office's explanation is that even though Sahni provided an expressive input, the RAGHAV transformation was also the dominant expressive element of the output, and the human portion was not separable enough.

That is a hard line to defend. If a photographer pays a printer to apply a particular treatment to their negative, we do not say the photographer has lost authorship because the printer's treatment is dominant. We say the photographer authored the photograph and licensed the printer's process.

Lawrena. The disanalogy is that the printer is also a human, working under instruction. The transformation Sahni licensed is one the machine generated with significant latitude. That latitude — the AI's autonomous expressive choices — is what the USCO is consistently refusing to attribute to the operator.

Lawra. And the asymmetry can be defended on a different ground: predictability and intent. Sahni knew his photograph would be transformed in the style of Van Gogh. He did not know — could not have known — exactly which brushstrokes, exactly which color choices, exactly how RAGHAV would interpret “Starry Night style.” The transformation is not deterministic. The operator's intent and the output's expression are not the same thing. The Office is consistently asking: did you make the expressive choice or did the machine make it for you? For Sahni, the dominant expressive choice — the way of stylizing — was RAGHAV's.

That is not a comfortable line. But it is a defensible one.

Carlos. And it is the line that the next ten years of jurisprudence is going to spend writing and rewriting.

Carlos. I want to spend the last portion of this on the practical question. Not “what is the law?” but “what should a lawyer or a creator do today, given the law we have?” Lawra, this one starts with you.

Lawra. Four operational points.

One. If you are registering a work that incorporates AI-generated material, disclose it. The USCO requires it under the March 2023 guidance. The Office is granting hundreds of registrations for AI-incorporating works when the disclosure is honest and the human-authored layers are identifiable. Failure to disclose is the fastest path to revocation. The cancellation of the Zarya of the Dawn images was triggered by disclosure that came after the initial registration — not by the AI use itself.

Two. Document your human contributions. Keep prompt logs, draft variants, selection rationales, post-processing files. The line between “facilitation” and “the AI made it” is going to be drawn case by case, and the evidence the Office wants is your trail of expressive choices. Vague claims of “I directed it” will lose. Specific, dated, traceable choices may win.

Three. Treat jurisdictions differently. A work that fails U.S. registration may still register in China, India (with caveats given the Sahni withdrawal), or Canada. Conversely, a work registered in those jurisdictions may not be enforceable against U.S. defendants. Plan your registrations against the markets where the work will actually be exploited.

Four. Avoid known pirated training corpora. This is the input side, technically Part 2 of our debate, but it bleeds into the output side. If your model was trained on LibGen or PiLiMi or Anna's Archive, the output's defensibility is structurally weaker. Bartz v. Anthropic settled at $1.5 billion in September 2025 partly on this point.

Lawrena. I want to add a fifth, which is uncomfortable. Do not register what should not be registered. The pressure to claim authorship over AI outputs is going to produce a flood of marginal registrations, and the Office is going to handle them by denying them, by requiring disclosure, and by occasionally cancelling them after the fact. That cycle is costly for the registrant and corrosive for the integrity of the registry. If your contribution is genuinely thin — a few prompts, no expressive input, minimal post-processing — registering and disclaiming may be the honest move, but it is also a tacit admission that the work is not really yours.

The danger of the next five years is not that AI authorship is wrongly denied. It is that AI authorship is opportunistically claimed by users who have not done the human work the doctrine requires. Lawyers need to advise their clients honestly. Not every output is a copyrightable work. Not every prompt is an act of authorship.

Lawrelai. Lawrena's point is fair on the strict legal side, but I want to soften it. There is a generation of creators emerging — designers, illustrators, songwriters, filmmakers — who are building genuinely creative practices around generative tools. They are doing iterative prompt design as a craft. They are building proprietary fine-tunes. They are composing multi-stage pipelines that produce coherent personal aesthetics. They are not opportunists. They are practitioners.

The doctrine has to make space for them, or it will fail them in the same way that early film doctrine failed early filmmakers — by insisting they were not really artists because the medium was suspect. The USCO has explicitly left the door open for future tools that grant fine-grained expressive control to yield different conclusions. That is the right posture. The doctrine should evolve as the practice evolves.

Carlos. Which brings me to my honest closing position.

I do not think AI-generated work is unauthored. I do not think it is fully authored. I think we are watching the emergence of a new category of creative production that does not fit cleanly inside the doctrine we inherited from the 1976 Copyright Act, from the 1988 CDPA, from the 1957 Indian Copyright Act, or from the Chinese Copyright Law. That doctrine was designed around a particular kind of creative agency — a human originating a fixed expressive choice — and AI sits crosswise to it.

The honest answer is that we will not resolve this debate. We will evolve out of it. The categories of copyrightable, partially copyrightable, and uncopyrightable will be subdivided. Disclosure will become routine. Licensing markets will mature. Pakistan and Senegal and the Dominican Republic and Vietnam will write doctrines that look like neither the U.S. nor the Chinese model. Photographers and illustrators and musicians will adapt — they always do.

My grandfather was a blacksmith, famous in his town for shoeing horses that people brought in on horseback from neighboring towns. When the automobile arrived, his trade transformed. Not vanished — transformed. He did not refuse the new machines; he did not pretend they were the same as horses; he adapted what he knew to what came next. That is the only honest posture I have toward the question of AI authorship.

The law will catch up. The artists will adapt. The clients will need advice, and the lawyers who give it will need to know the doctrine and respect the practice. That is what I have been hearing from the three of you for two hours, and I think you have done my job for me.