Someone at the firm still signs the work. The model does not.
If AI agents are already drafting the proposal, the memo, or the email a client will treat as yours, the rule is plain. Name the human who answers for the meaning. Say what the machine did. Do not let a detector score close the case.
Those are different decisions, and they get collapsed the moment a dispute starts. Disclosing a method is not proof of authorship. A detector score is not proof of cheating, or of innocence. Leave the distinction fuzzy, and the next argument will not sort it out for you.
He answers for pages he says he did not write
Thierry Ehrmann, founder of the art-market company Artprice, has tested the first half of that distinction in public. The company is promoting a free book, Dialogue Between a Thinker and AI, under a label that does not blink. In his preferred styling, it reads: “Human Thought & Dialogue: thierry | Writing: 100% AI.”
On Artprice’s account, he supplied the thought and the dialogue. The model supplied the prose. He is explicit about what he thinks that division means.
Writing is not merely holding a pen. It is answering for meaning. And I answer for the meaning of these pages.
He also argues that rejecting AI because it generates the text is a way of fetishizing manual labor. AI, in his phrase, is “the printing press of thought.” The spark, he says, stays in his brain. The model is a medium, not the source.
Artprice then states the limit the slogan can hide. Declaring the method does not legally establish authorship. A disclosure is not a copyright ruling. Get counsel before you treat a generated text as owned work.
That is the part of his claim a firm can actually use. He is offering responsibility. On his own company’s telling, he is not producing a title deed.
A flag is not a finding
A detector score fails the same test from the other direction.
On 6 October 2026, Simbarashe Kundizeza won the Caine Prize for African Writing for The Blue Fighter. The Guardian, reporting on 14 October, says that five days after the announcement Vauhini Vara published an Atlantic piece in which an editor had run the story through a detector called Pangram. Nearly the whole text came back flagged. The stated exception was an epigraph from an unpublished poem by the South African writer Vangile Gantsho.
Kundizeza’s reply to The Guardian was categorical: “I wrote ‘The Blue Fighter’. I did not use generative AI to write it. It is entirely my own work.” Treating a change of style, or a real improvement after five or more years, as proof of a machine, he said, “fundamentally misunderstands the craft of writing.”
Vara did not rest on the score alone. The Guardian reports that she cited stylistic habits: “not x, but y” constructions, dense metaphor, platitudinous aphorisms. She wrote, “Taste is subjective, but trust me, this story is bad, maudlin and full of tired tropes.”
As of that 14 October reporting, the prize had not withdrawn the award. It had opened a human review. Ellah Wakatama, chair of the Caine Prize board, and herself Zimbabwean, told Vara she wanted “proof that does not come from a machine.” If she asked him whether he was telling the truth and he said yes, she would believe him, “especially because the accusers are not coming from amongst our own people.”
On the BBC’s Front Row, in comments The Guardian reported, she said she was not against AI, but against its use as a final arbiter. She was “not happy for the work of writers, and in this case especially black writers, to form the testing ground for something that is at best still rough and ready.”
James Folta, writing in Literary Hub the same day, called that human-centered process the right call. Do not recruit him for a cheerful partnership between writer and model. His essay argues for holding the line against generative-AI experimentation. He treats the technology as a solvent on trust in prose, and he warns that even a careful AI-assisted publication can give cover to worse work. Citing him as a fan of transparent collaboration reverses the piece.
A score is not a verdict
The detectors in the 2023 research and the detector used on Kundizeza are different systems. Start there, or the numbers will get drafted into a case they do not decide.
Folta points to a 2023 Stanford Human-Centered AI write-up, “AI detectors biased against non-native English writers, ” on work by Weixin Liang and colleagues. The tools were older. They predate Pangram. They were described as near-perfect on essays by U.S.-born eighth-graders. They classified 61.22 percent of TOEFL essays, written by non-native English speakers, as AI-generated. All seven detectors unanimously flagged 18 of 91 of those essays. At least one detector flagged 89 of 91.
That is a serious false-positive problem in the systems that study measured. It is not a measurement of Pangram, and it is not a finding about The Blue Fighter.
Vara has said Pangram’s own research finds false positives rare, including for non-native English speakers. The Guardian also reports that she told the BBC there is relatively little third-party research. Folta writes that she emailed him that Pangram, “according to its own research as well as a small body of external research, is very effective.” Company studies plus a thin outside literature are not the 2023 result. The older bias evidence does not prove Pangram wrong. Pangram’s own studies do not prove it right.
Folta’s wider warning is the one a publisher, a law firm, or a sales desk should sit with. Once generated prose is common, doubt sticks to every text, and that doubt can be turned on writers of color. The opposite mistake is just as available: treat a flagged prize story as proof that a model writes as well as a person. He quotes former Caine winner Tope Folarin, via Vara, on the pressure writers from the global South feel to be “legible to the West, and obviously chatbots have glommed onto that.”
He also notes Vara’s earlier essay with GPT-3, “Ghosts, ” in The Believer, later reconsidered in Wired. The person who raised the flag has used the tool. That complicates the roles. It does not decide Kundizeza’s case, and it is not an endorsement of Ehrmann’s arrangement.
If a firm runs a detector at all, keep it as a tip. A smoke alarm is not a fire marshal, and a Pangram flag is not a verdict.
Same doubt, different outcomes
Set the Caine response next to the other two prize fights The Guardian records from 2026. The pattern is disagreement.
In July, Trinidadian writer Jamir Nazir won the Commonwealth Foundation short-story prize for The Serpent in the Grove despite social-media claims of AI markers. The Guardian does not report that the prize removed him.
In September, Canadian-Haitian novelist Thélyson Orélien was dropped from the Prix Goncourt longlist for C’était ça ou mourir. The Guardian says an anonymous online account alleged the book was almost entirely written by AI, and the title came off the longlist after that. The paper does not report that the jury ran a detector.
One prize left a winner in place amid social-media suspicion. One removed a book from a longlist after an anonymous allegation. One kept the award and opened a human review after a named detector and a published critique. A company that wants one “AI policy” to cover research help, rewriting, and signed prose is asking for a clarity these institutions did not have.
What the byline has to carry
Ehrmann’s label and the Caine flag fail in opposite directions. One tells you the method and still does not prove authorship. The other produces a score and still does not prove the method. A usable rule has to survive both failures.
Artprice’s market argument is narrower than the theory of the spark, and more usable. For ultra-contemporary work, the company says, value follows conceptual signature and command of the medium, not how elaborate the tool is. Output nobody guided, and output subordinated to an artist’s choices, should not be priced as the same thing. Clients already run a crude version of that test. They pay for judgment, not for keystrokes.
A percentage will not write the policy. Ten, twenty, or thirty percent generated prose (figures Artprice offers as illustrations, not as a count of this book) do not show where a thought started, or who answers if the result is wrong, derivative, or merely fashionable.
Split the work by what the firm is willing to sign. This is the practical core of AI automation, and of AI for business, once a human name is on the page.
- Research and outlining. The system gathered or arranged material. Whoever signs still has to stand behind the facts. If a reader would assume the file was built by hand, say a tool was used to search or structure.
- Rewriting. The firm’s argument may survive. The sentences may be the model’s. Those are different facts, and the note should separate them.
- A client-facing paragraph the firm will sign. This is the line that matters, including in AI for sales, where the outreach is the product. The signer answers for the meaning. If it is wrong or derivative, the disclosure does not move that duty onto the software.
The note has to be dull enough to audit later. Who directed it. What was accepted unchanged. Who will defend it if it fails. “AI-assisted” does not say any of those things.
None of that settles a separate fight: whether the model was allowed to learn from other people’s books, voices, or likenesses. A byline that admits automation is not a receipt for the training set. Keep that limit visible. Fold it into the authorship label, and both questions become impossible to manage.
Questions worth asking out loud
- Did a detector decide the Caine Prize?
Not in the reporting through 14 October 2026. The Guardian and Literary Hub both describe a Pangram flag on nearly all of The Blue Fighter, a human review, and an award left in place. Kundizeza says the story is entirely his.
- Did the Goncourt jury catch Orélien with a machine?
That is not established. The Guardian reports that he was removed from the longlist after an anonymous allegation of near-total AI authorship. It does not report that the jury ran a detector.
- Do the 2023 Stanford numbers indict every detector?
No. Liang and colleagues, as cited by Folta from the Stanford Human-Centered AI write-up, documented severe false positives from earlier tools on TOEFL essays by non-native writers. Those figures do not measure Pangram, the system used on the Caine story.
- Does “Writing: 100% AI” make Ehrmann the legal author?
Artprice says the transparency does not legally establish authorship. The label is a claim of responsibility, not a copyright ruling.
- What should the byline disclose, and what does that not settle?
Name who directed the work, what the machine did, what was accepted unchanged, and who will defend it if it is wrong or derivative. That does not settle copyright, permission to train on other people’s work, or whether a detector was right.
The prizes and the clients will still want a name attached to the meaning. Put that name on the work before a detector tries to supply one for you.