In March 2026 a publisher cancelled the US release of a horror novel and withdrew the UK edition after allegations that AI had written parts of it. The author says someone she hired to work on an earlier self-published version used AI without her knowledge.
Now hold that next to a ghostwriting contract. The author was held responsible for what a person she paid did with a keyboard she never watched. That’s the risk you carry the moment you hire anyone to write under your name, and it existed before any of this. AI just made it cheap.
What is happening to authors right now
By August 2026, reporting counted at least three major publishing agreements collapsing over AI allegations in a matter of months, with no wrongdoing established in any of them. The accusation did the damage on its own.
The 2026 Commonwealth Short Story Prize is the case worth studying, because it went the other way. One of the winning stories, published in Granta, scored 100 percent AI-generated on a widely used detector after readers ran it through. The author said his process uses the text-to-speech function on his phone. The Commonwealth Foundation reviewed working drafts, time-stamped documents and notes, found no evidence of AI use, and stood by the story.
Look at what settled it, because it is available to you too. Not a rebuttal, not a better detector, not a statement of good character. Drafts with dates on them.
Meanwhile the tools to accuse are now one tap from every reader you have. A detector is embedded in a major publishing platform. Another platform added a button letting users report posts as AI slop. None of it arrives with a standard of proof, and none of it was designed to survive being wrong.
How reliable are AI detectors on a finished manuscript?
The vendor of the best-known detector claims a false-positive rate of 0.01 percent. A researcher who studies these systems makes the point that even one percent is unacceptable in publishing, where an accusation by itself ends careers.
Run the arithmetic against your own submissions. At one percent, a publisher processing ten thousand submissions falsely accuses a hundred writers. At 0.01 percent it’s one, assuming the vendor’s number holds outside the lab, and lab numbers rarely do.
The failures are not random either. A 2023 Stanford study found detectors biased against non-native English speakers, whose writing draws on a smaller pool of constructions and reads as more predictable. One writer ran a 2010 atmospheric physics thesis, submitted more than a decade before this technology reached the public, and got roughly 70 percent. Formal, careful, low-surprise prose is exactly what these tools flag.
Which means the writers most likely to be falsely accused are the disciplined ones, the technical ones, and the ones writing in a second language. I’ve written separately about why a detector result cannot be evidence, and the short version is that it measures resemblance, never authorship.
Which four questions settle it?
Start by not opening with the number when you raise it. Once you say “the detector says 94 percent,” the conversation is about the tool, and you’ll spend a week arguing about software instead of finding out what you hired.
Ask for the trail. Four questions, and they take one email.
Can I hear the interview recordings this chapter came from? A ghostwriter working properly has hours of you talking, and you should be able to hear them. If the recordings exist, you can hear your own stories in your own voice, told before anything was written.
Can I see the outline I approved, and the draft that followed it? The outline predates the prose, and you signed off on it. A book built from an approved plan leaves that plan behind.
Can I see the version history with dates? Chapters delivered one at a time, your comments on them, the next version with those comments applied. That sequence is the fingerprint of somebody working for you, and no prompt produces it.
Walk me through why this chapter opens the way it does. This one does more than the other three combined. A writer who built the chapter will tell you why the second scene comes before the third, what they cut and why they cut it. Someone who generated it can’t, because the reasoning never happened. You’ll know inside five minutes.
That’s it. Four questions, no software, no accusation, and you get a clear answer either way.
What does an honest process leave behind?
I’m asking you to trust a different kind of evidence, so here’s mine, in the order it accumulates.
Recorded interviews first, hours of them, with the client doing most of the talking. Transcripts of those recordings. An outline built from the transcripts, sent to the client and approved before a word of prose exists. Chapters written one at a time, each delivered for comment. The client’s comments. The revised chapter showing them applied. Dated files going back to the first conversation.
None of that was built as a defense. It’s what doing the work properly looks like, and it has been my process since long before anyone worried about machines. That’s the reassuring part: the protection is a by-product of ordinary competence. You don’t need a writer with a clever policy. You need one who works in a way that leaves a record.
The interview process itself is the thing producing both the book and the proof.
Why a contract clause banning detectors is the wrong answer
Some writers have started putting language in their contracts saying that if a client runs the manuscript through a detector, the project ends.
I understand the impulse. Being accused after months of honest work is infuriating, and the tool doing the accusing is unreliable. But put yourself on the other side of that page. You’re about to pay someone tens of thousands of dollars, and the contract says checking is grounds for termination. What would you think?
I’d think the writer was worried about what a check might show. It doesn’t matter whether that’s fair. The clause creates the suspicion it was written to prevent.
What does belong in the contract is the opposite: a statement of how AI is and isn’t used on the project, agreed before the work starts. Industry people who track this expect AI language to become as routine as rights and royalties clauses, and disputes and lawsuits between authors and writers over AI use to follow. The protection is agreeing the rules in advance, not forbidding the question.
The reverse problem nobody warns authors about
One more, because it’s the version most likely to catch you.
The publisher pulled that horror novel over work done by someone the author hired for an earlier edition. She didn’t write the AI text. She may not have known it existed. Her name was on the cover, so it was her problem.
If you hire a ghostwriter, an editor, a developmental reader or a book doctor, whatever they produce becomes yours the moment it goes in. That’s the actual exposure, and no detector protects you from it. What protects you is knowing how your writer works before you pay, asking the four questions along the way instead of at the end, and keeping what they send you so a dated trail exists on your side too.
What this is really about
Strip away the software and you’re left with an old question. Did the person you paid do the work?
Publishing has always had that question, long before you or I got here. Manuscripts have been farmed out, padded, and written by someone other than the name on the spine for as long as there’s been a trade. What changed is that the farming out got cheap and a tool appeared that claims to detect it and can’t.
So we’re back to evidence, where we were before anyone built a classifier. The recordings, the drafts, the dates, and a writer who can explain their own chapter. Ask for those. If they’re there, you have your answer. If your writer can’t produce a single file older than the final one, the detector didn’t tell you that. The empty folder did.
For the wider picture of AI and books without the panic, my AI writing hub collects it. If you want a book with a working trail behind it from the first recorded interview, that’s what my ghostwriting service produces.