.png)
AI Detectors Have an Authorship Problem. Writers Need Provenance, Not Probability Scores
AI & The Author — 28 September 2026
The argument about artificial intelligence and books is moving into a more uncomfortable phase. It is no longer only about whether a machine can produce convincing prose. Increasingly, writers are being asked to prove that they did not use one.
In this article
• Why AI-detection claims are becoming an authorship problem
• The difference between AI-assisted work and AI-authored fiction
• Why provenance and disclosure are stronger than probability scores
• What authors and publishers should do next
The detector has become part of the story
Fresh reporting today on the controversy surrounding Canadian-Haitian novelist Thélyson Orélien has put AI detection itself under the spotlight. The debate matters beyond one author or one book. Once a detector score is treated as evidence of authorship, a statistical tool can begin to function like a verdict — even when the underlying question is far more complicated.
That is a dangerous shortcut. A novel is not simply a block of text waiting to be classified as human or machine. It is research, planning, drafting, rewriting, editorial judgement, discarded scenes, notes, conversations, corrections and thousands of decisions that never appear on the final page. A detector sees the final text. It does not see the creative history that produced it.
The current controversy is summarised in Agence France-Presse reporting published today, which focuses specifically on the reliability of AI-writing detectors.
Publishing already uses AI. just not always where readers see it
There is another reason the debate needs more precision. AI is already inside publishing. It is being used for metadata, proofreading, customer service, royalty administration and other back-office tasks. Publishers Weekly recently described an industry still wrestling with AI while also reporting substantial organisational adoption.
That broader industry picture is useful context: Publishers Weekly's Publishing's AI Reckoning describes AI becoming part of everyday publishing operations rather than remaining an experimental curiosity.
For authors, this distinction matters. Using a tool to organise research, test an outline, check continuity, visualise a setting or help prepare marketing material is not the same creative act as asking a system to generate a novel and putting your name on the result. Collapsing all of those activities into the phrase 'used AI' makes the discussion less useful, not more.
Assistance is not authorship
My own line remains straightforward. The books are mine. The characters, worlds, story decisions, prose and final editorial judgement belong to the author. AI can be useful around the creative process without becoming the creator of the work. That is why transparency matters — but it has to be meaningful transparency, not a ritual confession that treats every use of software as equivalent.
Writer and publishing commentator Jane Friedman's regularly updated AI and Publishing FAQ makes a practical distinction when discussing research, brainstorming, outlining, editing and copyright.
The better question is therefore not 'Did AI touch this project?' It is: who made the creative decisions, who is accountable for the words, and what role did the technology actually play? Those questions tell readers far more than a detector percentage ever could.
Copyright is pushing the same argument in another direction
At the same time, authors are still fighting over what happens to their work before an AI system ever produces an answer. In July, a US judge approved Anthropic's $1.5 billion settlement of a copyright lawsuit brought by authors who accused the company of misusing books to train Claude. The settlement does not resolve every legal question around training, but it demonstrates how valuable provenance, licensing and lawful acquisition have become.
For the underlying settlement, see Reuters' report on the court approval.
More recently, disputes have continued over who is entitled to settlement payments for individual books. That is another reminder that the AI argument is increasingly about records: who owned what, when rights reverted, what material was used, and under what authority.
The continuing allocation dispute is covered by TechCrunch.
Provenance beats detection
That suggests a more useful direction for publishing. Instead of trying to infer a book's creative history from the statistical texture of its sentences, build a clearer record of the process. Manuscript versions, revision history, editorial correspondence, source notes and sensible disclosure can provide positive evidence of authorship. They show what happened rather than guessing from the finished prose.
This does not mean every writer should be required to surrender private drafts or document every keystroke. Nor should authors be presumed guilty because their prose happens to resemble a detector's idea of machine-generated language. The principle is simpler: when trust matters, evidence of process is stronger than automated suspicion.
Publishers also need consistent rules. If a house uses AI internally for metadata, editing support or marketing, it becomes difficult to defend vague contractual language that treats any author use of AI as inherently suspect. The industry needs definitions that distinguish generation from assistance and authorship from tooling.
What I would like to see next
First, publishers should state clearly what uses of AI require disclosure and why. Second, AI-detection scores should never be treated as standalone proof of machine authorship. Third, authors should retain enough of their normal working material to establish provenance if a serious dispute arises. And finally, AI companies should face the same transparency pressure that creators increasingly face: where did the training material come from, was it lawfully obtained, and how are rightsholders treated?
The technology is not going away. Neither is the anxiety around it. But we will make better decisions if we stop treating 'AI' as one undifferentiated thing. A tool used around a human creative process is not automatically the author of that process and a probability score is not automatically proof that the human disappeared.
The question for writers
If a publisher challenged your manuscript tomorrow because an AI detector flagged it, what evidence would you expect them to produce and what evidence of your own creative process would you be comfortable providing?
Sources & further reading
About Rob Frankson
Rob Frankson writes science fiction and publishes 121 Minutes, exploring storytelling, technology and the practical realities of being an independent author. About Rob Frankson.
AI & Editorial Transparency
AI & The Author is edited and published by Rob Frankson. Artificial intelligence is used to assist with news research, initial drafting, content organisation and supporting imagery. All articles are reviewed and, where necessary, edited by Rob Frankson before publication. The opinions, editorial position and final decision to publish remain the author's.
Comments