Curation Is Craft the Copyright Office Can See
Translation, code ports, prompt-select-edit loops. Same old logic, new source text.
Previously: Channeling, Translation, and the Urantia Book.
When a translator takes a poem from Greek into English, no one seriously argues that the translation belongs to Homer. The translator made thousands of decisions — which English word carries the ambiguity of the original, where to break a line, whether “wine-dark sea” stays or becomes something more literal, something less. Those decisions are original expression. They’re protectable. The translator owns what the translator made, even though the source material isn’t theirs and the final text wouldn’t exist without it.
That logic has been sitting in copyright doctrine for a long time. It applies to code ports, to film adaptations, to scholarly editions with original apparatus. It quietly governs what happens when a human being assembles a work from AI-generated material, too. The law isn’t starting from scratch here. It’s applying an old principle to a new kind of source text.
The Urantia lesson, one step further
A note on sourcing: the legal analysis in this section and the next draws on a case brief rather than the primary opinions; for legal purposes, go to the originals.
Last week I wrote about how the Contact Commission of the Urantia Foundation — the group that curated and structured the claimed spirit dictations — ended up being the reason the Ninth Circuit upheld the copyright in Urantia Foundation v. Maaherra, 114 F.3d 955 (9th Cir. 1997). The court found, as the brief I worked from characterizes it, that the Commission’s “selection, coordination, and arrangement” of the revelations into papers, their choice of which questions to ask and in what order, constituted the human creative spark the law requires. The spirit beings — actual or imagined — wrote nothing protectable. The humans who shaped and organized the material did.
The implication runs in two directions. Passive transcription of a non-human source isn’t authorship. Active curation of that source — decisions about structure, sequence, what to include and what to leave out — is exactly what authorship can look like, even when you didn’t generate the raw material yourself.
That’s the opening the U.S. Copyright Office’s guidance on AI-assisted work walks through — the 2023 guidance and the 2025 report both address this territory, as commentators have noted. The Office has said, consistently, that raw AI output without human selection or arrangement lacks protection. But human-authored expression created with AI assistance retains its copyright. And “creative selection, coordination, arrangement, or modification” of AI-generated material may qualify on a case-by-case basis. The question isn’t whether AI touched the work. Whether a human being exercised meaningful creative control over what survived — that’s what matters.
The Thaler boundary
Stephen Thaler asked a federal court to register a copyright for a visual artwork, “A Recent Entrance to Paradise,” generated autonomously by his AI system, the “Creativity Machine,” with the machine listed as sole author. That case, Thaler v. Perlmutter, reached the D.C. Circuit and was decided, as the brief I read places it, in 2025. He argued the machine did the creating; he owned the machine; therefore he should own the output under work-for-hire doctrine.
The D.C. Circuit said no. The Copyright Act assumes authors are human. The court pointed to provisions in the statute about human lifespans, widows, heirs — the machinery of a law written for people, not systems. A machine can’t be an author under that framework, and claiming it is doesn’t transfer authorship to its owner any more than owning a camera meant Ansel Adams owned the light.
The Thaler case matters because it draws the boundary at the clearest possible point: a work generated entirely autonomously, with no human selection or arrangement of the output, sits outside copyright protection. What that boundary does is clarify, by contrast, where protection begins. It begins when a human being makes choices.
What the brief literature is beginning to take seriously — and this may be secondary commentary rather than the court’s own reasoning — is that if Thaler had framed his contribution around his own role in guiding, selecting, and arranging outputs, the analysis might have been different. That’s not a gap in the law. It’s the same structure translation copyright, compilation copyright, and derivative-work copyright have always occupied.
What counts as substantial human control
Here is where the analysis has to be honest rather than wishful.
Typing a prompt and accepting the first output isn’t curation in any meaningful sense. It’s closer to what the Urantia district court called passive transcription — a person as a mechanical conduit, disclaimed as a creative actor. The Copyright Office’s 2025 guidance says the same thing: prompts alone, given current technology, don’t give the user sufficient control over the expressive elements of output to ground a copyright claim — at least as the report is described by commentators who’ve read it closely.
But a prompt is rarely just a prompt in a serious working process. What actually happens in a loop of AI-assisted writing, at least in mine, looks more like this: a first draft comes back from the model. Most of it is wrong — wrong tone, wrong angle, wrong word choices. I rewrite sentences. I move paragraphs. I cut an entire section and reroute the argument. I prompt again with tighter constraints, not because I liked the direction of the first pass but because I want to see whether the material I’m trying to get at will emerge differently under new pressure. I keep some things and throw out others. A structure takes shape that has more of me in it than of the model.
That is selection, coordination, and arrangement. It’s also, candidly, closer to what Ted Chiang means when he argues that art is the accumulation of choices — that every scale of decision, from argument structure to sentence rhythm to individual word, is where meaning lives. Whether AI helped isn’t the question. Whether the human made enough of the decisive choices to constitute an original expression — that’s what the copyright analysis turns on.
The Copyright Office thinks this can qualify. It has to be demonstrated case by case, not assumed. The more the workflow resembles passive acceptance of model output, the weaker the claim. The more it resembles a translator’s persistent negotiation with a source — refusing defaults, rewriting, selecting what serves the work’s purpose — the stronger it is.
I’m not a lawyer, and nothing here is legal advice.
The assembled work distinction
There’s a useful distinction between a raw AI output and an assembled work.
A raw output is what the model generates when you prompt it once, accept it, and move on. It may be competent. It may even be good. But it carries the statistical signature of pattern completion: default structures, default rhythms, default resolutions. In the language I used in the second post in this series, it’s a reflection from a hall of mirrors — and if you accept it unchanged, you’re accepting the distortions as your own.
An assembled work is what happens when a human being treats that output as material rather than product. The prompt-select-edit loop, when it’s actually working, is less like ordering from a menu and more like milling rough lumber. You end up with something that started as raw output but has been cut, shaped, and fitted to a specific purpose. The evidence of your choices is distributed through the work.
This is why the compilation framework is useful beyond its legal implications. Legally, it describes what kind of human contribution can generate a copyright claim. As a description of creative labor, it maps onto what craft means in the context of AI-assisted work: not the absence of the tool, but the presence of the maker’s judgment in every choice that mattered.
The ceramics analogy I’ve been using in this series holds here too. A potter who buys a bag of commercial clay and a jar of store-bought glaze is still a potter. The craft question is what they do with those materials — whether the form they make is theirs, whether the decisions that shaped it came from somewhere real. Buying the materials doesn’t answer that question. Neither does generating the first draft.
What curation actually asks of you
Curation, in this sense, isn’t a passive or managerial role. It’s an active creative stance that requires you to know what you want well enough to recognize when the model hasn’t given it to you — and to care enough to refuse the default.
That refusal is exactly what the Copyright Office is looking for when it asks about substantial human control. It’s what the Ninth Circuit was describing when it praised the Contact Commission for shaping the Urantia questions rather than merely receiving dictation. A translator makes the same choice when she decides that a literal rendering loses the poem’s force.
The law can’t see into your process. It can only see the work. But the work is where your decisions either show up or they don’t. If they show up — if the structure, the selection, the sentence-level choices are yours — then you’re not just a user of the tool. You’re its author in the only sense of that word that matters.