Jerry Falade’s own agents took his book away from him. His debut crime novel drew a two-book offer from Minotaur reported at more than two million dollars, people who had read the manuscript started saying it looked like AI, and within days his own representatives withdrew it, because they could no longer establish how it had been written.
In this post I will:
Show you exactly what the evidence in this case was.
Give you the audit trail to build, and the two steps that matter if you only have twenty minutes.
Say plainly how I use AI to write this newsletter, including the bit where a detector calls me a machine.
Most of you are writing a newsletter, or a dissertation, or a report your manager will read on a train. The mechanism that ended Jerry Falade’s deal is the one now sitting under all of it.
Call Me, I’ll Hide the Body went to a fourteen-way auction and drew that offer from Minotaur, an imprint of Macmillan US. Reports of the figure vary from two to two and a half million. His co-agents at Europa Content, Marc Gerald and Ashley Coleman, pulled it.
Falade denies using AI.
“The accusations are wrong, and I am completely innocent,”
he told the reporter Jeff Sneider. Adding:
“I was on the verge of success, and then all of a sudden, there were rumors, but no one asked me anything. I don’t know why my agency did not have my back.”
Accusation is now one tap away
As I have previously discussed Substack partnered with the detector Pangram on 21 July 2026. Any reader on web or iOS can press a button on an eligible post and get a number back about whether a human wrote it.
LinkedIn have now also added a button that lets users report posts as AI-generated slop, on a platform that spent two years encouraging everyone to generate posts using AI. Your university has a detector. Your publisher has a policy.
The tools to accuse you are one tap from every reader you have, and none of them come with a standard of proof.
The evidence was a book review
The case against Falade’s manuscript, as reported, was three things:
Negative parallelism, meaning sentences built on ‘this, not that’.
Off-kilter metaphors.
A flat prose style.
Those are criticisms. They belong in a workshop, or a two-star Goodreads review, or the margin of a first draft. Every one of them describes a large quantity of published human writing, some of it very good.
There is a second strand. The Bookseller reports a meeting on 29 July at which aspects of Falade’s account changed, and that this is what triggered the withdrawal. However, the agency has so far produced no detector log, no set of drafts, no version history, and no timestamp.
Sandy Hodgman of Hodgman Literary, who handled the potential foreign and UK rights, put it like this:
“Unfortunately, we are no longer able to authenticate how the manuscript fully evolved from origin to completion.”
They had accepted his assurances at first. Then they could not substantiate them, so the book was withdrawn.
Who is being asked to prove themselves
This is already happening at scale. Hachette cancelled the US release of Shy Girl by Mia Ballard in March 2026 and withdrew the UK edition, after a lengthy investigation prompted by evidence the New York Times brought to it. Ballard says an acquaintance she hired to work on an earlier self-published version used AI without her knowledge, and she has told the New York Times she is pursuing legal action.
Falade is a young Black writer and a doctoral student. He has said publicly that three Black authors landed major deals this year and all three saw those deals cancelled or disrupted after AI suspicion.
As I have written about numerous times now, researchers at Stanford, writing in 2023, ran seven widely used detectors over 91 essays written by non-native English speakers. Across the seven, the average false-positive rate on that human writing was 61.3%, and 97.8% of the essays were flagged as AI by at least one detector. The same detectors read US eighth-grade essays with near-perfect accuracy. Careful, standard, unadorned prose looks machine-made to many of these AI detectors, and that describes most people writing in a second language and most people taught to write formally.
The error in the Stanford study fell overwhelmingly on one group of writers. When the tool is uneven, asking who keeps getting called out is a reasonable question.
What produced the verdict in the Falade case was rumour, a changed story, and a reading of the prose, with no record on either side.
How to build an audit trail, starting today
An audit trail is a record of how a piece of writing came to exist, made while you are writing it.
Here is how to create an audit trail for your own work. If you have twenty minutes, do 1 and 4. Version history and an AI log cover most of what anyone will ever ask you for.
Write somewhere that keeps its own history. Google Docs keeps full version history for free, and so does Word with AutoSave on. Substack’s own editor keeps drafts and revision timestamps, so if you draft in the app you already have more than you think. Version history is the single strongest artefact you can hold, because the timestamps are set as you go. You can’t manufacture three weeks of edits after an accusation arrives.
Keep the ugly early drafts. Do not overwrite. Save dated copies at real milestones: the first mess, the structural rewrite, the version you sent to a friend. Five saved drafts across three weeks say more than any detector output ever will.
Keep your raw input, whatever form it takes. This might be the photograph of a notebook page, a voice memo in the car, a scribbled outline, or the twenty-eight-tab research session you had open. Keep it.
Log the AI, specifically. If you use AI in your writing process than keep a single running file per project. One line per session, copy this shape:
2 Aug | Claude Opus 5 | structural edit, section 3 | took the reordering | rejected the new opening
Date, tool, what you asked for, what you took, what you refused. This is the artefact almost nobody has, and it is the one that answers the question Falade’s agents actually asked.
5. Ask your editors and contractors what they use. Ballard’s account of her own case turns entirely on this: work she paid someone else to do, using tools she says she did not know about. If you hire a developmental editor, a copyeditor, a ghostwriter, a VA, or a cover designer, put one line in the agreement asking them to disclose AI use and to keep their own drafts. You are responsible for work that goes out under your name, including the parts you did not do.
6. Keep your research trail. Sources, links, saved PDFs, and notes on what you read and when. A piece that can name where every claim came from reads as researched, because it was.
7. Write your AI use statement before anyone asks for it. Three or four sentences on how you work, published somewhere durable: an about page, a pinned post, the back matter of a book. Here is a template to adapt:
I use [tool] for [specific tasks: research scanning, structural feedback, proofreading, and image generation]. I do not use it to [generate first drafts / write in my voice / produce the arguments].
An accurate statement about heavy AI use is worth more than a flattering one that falls apart.
8. Put a provenance clause in the contract. For book deals and commissioned work, agree upfront what evidence you would provide, to whom, on what timescale. Publishing contracts routinely carry an AI warranty now. Very few of them define what proof would look like, or name who decides. Get that written down while everyone still likes each other.
This is not a guide to evading detection
Some people will read those eight steps as advice on how to look human.
Writing to beat a detector means changing your sentences to fool a classifier. It makes your prose worse, it is a losing game against a system that updates without telling you, and it is dishonest.
Building an audit trail means keeping a record of what you did. It changes nothing about your writing. It works whether you use AI heavily, lightly, or not at all, because the record simply says what happened.
If you have never used AI and you kept nothing, you are as exposed as anyone else here, and that is the unfairness at the centre of this. Your defence is a record, and you deserved to be believed without one. Unfortunately many will not.
If you enjoy this newsletter then you might also enjoy Slow AI the book.
My own trail, since I am asking for yours
Pangram scores my writing as 100% AI generated. The maximum the system gives. I wrote about the rollout in Substack’s AI Detector and the Return of the Witch Hunt, where that post scored 100% machine and a humanised version of the same text scored 100% human.
My writing is AI assisted, and I have never once claimed otherwise. So the score describes the surface of the text. It says nothing about how the text was made, and that is what anyone actually wants to know.
Here is mine, written out, because I am asking you to write out yours.
Finding the subject. I watch what my own readers argue about in the notes, and I run agent checks across social platforms and news feeds to see what is surfacing and what is already exhausted. The machine does the scanning. I decide what matters this week from what comes back.
Drafting. I mostly do this via Whisper Flow and I speak the draft out loud, making edits as I go, combing it with snippets of notes that I have taken during the week via the research phase.
Editing. I hand the draft to Claude and ask it to peer review, to argue with the structure, and to suggest edits. I also run a persona check, five imagined readers from a near non-user to a hostile expert, to find where the piece loses people. I take some of it and ignore plenty.
Everything around the words. The SEO, metadata, and artwork are all generated using AI.
Then there is my thesis
Take something you wrote before ChatGPT existed and put it through a detector. Use something old, where you already know the answer.
I ran my PhD thesis. Atmospheric physics, submitted in 2010, at a point when the technology being accused of writing it was more than a decade from public release. Running it through a detector in 2026 returns around 70%.
A physics thesis is exactly the careful, formal, low-perplexity register that the Stanford study found classifiers flag.
A detector tells you one thing: that a piece of text resembles a statistical distribution. The question in the room is who wrote it, and no detector has ever answered that one.
Why this matters
None of this should be your job. You should be able to write a book, sell it, and have the people who represent you assume you wrote it. That world is gone, and I am not going to insult you by pretending it comes back if we are patient.
Two years ago the burden of proof sat with the accuser. It has moved. As I wrote in Guilty Until Proved Human, we now start from suspicion and work backwards, and the person carrying the cost is the one being asked to prove a negative, which cannot be done. What you can do is make the question answerable.
Falade lost a two million dollar deal in the gap between an assurance and a record. Fill your gap this week. Turn on version history, start the log, and write the four sentences about how you work.
As you do this, notice who gets asked to prove the provenance of their writing, and who never gets asked at all.
What is your process? Tell me in the comments how you actually write, AI or no AI. There will be no judgement here, as I would like this thread to be a safe space where we can learn from each other as readers, and writers, and humans.
Go slow.
Why don’t we start requiring authors to highlight which lines of dialogue are theirs and which ones were modified by a human editor? We don’t do that because it’s fucking absurd.
I'd be interested in a statement from the publisher as to their objection to minor use of AI. If their stance is that it's not the author''s own work, I don't see how that is defensible. Using Grammarly means it's not all the author's own work. Ditto using a copy editor. Ditto asking friends to read a draft and give feedback. And using a ghost writer?
Maybe it's time for Arthur Miller to rewrite The Crucible, replacing witchcraft with AI.