Applied Case: The Copy Accuses the Source
When Modal Path Ethics first examined the collapse of Jerry Falade’s extraordinary publishing deal for Call Me, I’ll Hide the Body, I refused to decide what the public record could not decide.
There were three possible fields.
- Falade wrote the book.
- Falade substantially used generated prose and denied it.
- Falade used artificial intelligence somewhere inside the large middle region between ordinary software and synthetic authorship.
Every one required a different ruling. So the first article stayed there.
Authorship is a path claim.
The manuscript had passed. The path became the case.
Now, Falade has described the path himself.
In an essay published by Kathleen Schmidt on September 17, Falade says he wrote Call Me, I’ll Hide the Body.
Kathleen Schmidt
He describes the Yorùbá oral traditions informing its voice, the drafting artifacts he still possesses, the limited research purposes for which he says he used a locally hosted language model, the accusation that followed him through publishing, and the process by which an unpublished manuscript was submitted to AI-detection software without his knowledge.
He says he has handwritten material, messages, multiple drafts, and Google Docs history capable of showing the manuscript developing over time.
This article is going to grant him his account.
That makes the original problem even worse.
Because under Falade’s account, publishing did not discover that artificial intelligence had impersonated a human writer.
It encountered a human writer whose writing resembled patterns machines had learned from humans.
Then the copy accused the source.
The Copy Accused the Source.
Falade’s most important contribution is cultural.
He says the voice of Call Me, I’ll Hide the Body carries Yorùbá storytelling habits: proverbs, aphoristic expression, metaphor, parallel structure, verbal rhythm, and emphatic contrast.
Then he discovered that several of those features had entered the growing folk taxonomy of "AI tells."
Jane Friedman later reported that agents who had considered the manuscript described its opening as containing “hallmarks of AI writing,” while also reporting that Falade’s former agents had not themselves relied on AI detection.
Falade’s answer is devastatingly simple.
- Those habits came from somewhere.
- They came from people.
He describes learning aphoristic and metaphorical language through Yorùbá culture and literature long before generative artificial intelligence existed. He even went back through his approximately 85,000-word manuscript and counted the stylistic features now being treated as suspicious: eighteen instances of the famous “not X, but Y” construction and forty-six em dashes.
The man began conducting forensic analysis on his own punctuation.
The deeper problem is structural.
- A language model learns statistical regularities from human language.
- Those regularities become recognizable features of model output.
- Observers begin associating the features with artificial intelligence.
- Then a human being uses one of those features.
- Observers begin associating the features with artificial intelligence.
The resemblance travels backward.
Machine imitation becomes evidence against the thing being imitated.
At best, that inference identifies resemblance. It cannot possibly authenticate provenance. And resemblance is especially dangerous where the reader has weak prior contact with the human tradition producing the prose.
- A familiar American rhetorical habit may register as style.
- An unfamiliar Yorùbá-inflected habit may register as anomaly.
And once “anomaly” and “AI” become neighbors in the evaluator’s incredible mind, cultural distance can start impersonating technical evidence.
The detector does not even have to make that mistake by itself.
People can do it first.
That was already one of the warnings in the original Modal Path Ethics article: literary judgment cannot be promoted into forensic machinery simply because a new forensic anxiety has appeared.
Falade’s account gives that warning a much more specific body.
Can a detector distinguish human from machine?
Is no longer sufficient.
There is another question underneath it:
Which humans have already been placed closer to “machine” by this evaluator’s expectations about what human writing is supposed to sound like?
The Detector Becomes an Editor.
Falade describes writers becoming afraid of parallelism and other stylistic devices because they may trigger suspicion. He recounts another writer intentionally introducing errors in an attempt to look more convincingly human.
That changes the function of the detector.
The detector is no longer only evaluating prose after the prose exists.
The detector is changing what prose people are willing to write.
- A writer reaches for a rhythm.
- Stops.
- Too AI.
- Stops.
- A writer builds a symmetrical paragraph.
- Fucking breaks it.
- Too AI.
- Fucking breaks it.
- A writer likes em dashes.
- The machines got custody of those.
- A multilingual writer reaches naturally toward the syntax carried from another language and now has a second question riding beside the sentence:
- Will this look human enough to a machine built to recognize machine writing?
At that point, the "detector" has become an editor through threat.
That machine is selecting style without ever opening the manuscript.
And the selection pressure is grotesque.
- Write cleanly enough to be published.
- Write strangely enough to be believed.
The more writers adapt themselves to the detector, the less the detector is observing an independent human-writing distribution. It is helping produce the distribution it will later measure.
In order to "protect human authorship from AI," publishing is using AI to change how human authors write without AI.
This is exactly the kind of active-information problem Modal Path Ethics keeps finding in instruments that are supposed to observe fields from outside them.
- The instrument enters the field.
- The field reorganizes around the instrument.
- Then the instrument measures the world it helped make.
- The field reorganizes around the instrument.
Research Has a Jurisdiction.
Falade also directly answers the most badly designed question in this entire controversy.
DID YOU USE AI?
According to his essay:
- Yes.
- For research.
Falade says he used a locally hosted model, Dolphin-Phi, while writing a crime novel whose protagonist has chemistry expertise. He describes using it for questions involving forensic and crime-fiction research and for converting measurements into American units. He says the novel’s prose remained his own.
That is almost comically close to the taxonomy the first Modal Path Ethics article demanded before Falade published this account.
- Research is one transition.
- Generated expressive language is another.
- Retaining generated language is another.
- Structural assistance is another.
- Disclosure is yet another.
- Candor about disclosure is another still.
The question “DID YOU USE AI?” compresses all of them into one bit.
Under Falade’s account, that bit returns a YES.
Meanwhile, the question everybody actually fucking cares about
Who determined the expressive language of Call Me, I’ll Hide the Body?
returns FALADE.
Turns out those are completely compatible answers.
This is why contamination logic is so useless.
A novelist can use an artificial-intelligence system to investigate whether something in a crime scene is physically plausible without transferring authorship of the scene to the machine.
The distinction is ordinary everywhere else.
- A novelist who calls a pathologist has not coauthored the novel with the pathologist.
- A novelist who consults a conversion table has not transferred expressive control to Celsius.
- A novelist who searches an archive has not converted the archive into the narrator.
The instrument matters according to what transition it performs.
Under Falade’s account, artificial intelligence entered the research path.
That does not make it the author.
The Audit Entered the Manuscript.
Then comes the strangest part.
Falade says a journalist obtained his unpublished manuscript, submitted it to AI-detection software without his knowledge or consent, and sent the resulting report to his creative partners.
Schmidt identifies the detector as Pangram. Of course it was fucking Pangram.
Publishers Weekly had already reported in August that some industry publications had run the manuscript through Pangram, while noting that the precise basis on which the book had been judged remained unclear.
The provenance investigation had now become a provenance event.
That is not a small detail at all.
Investigative reporting cannot depend on the subject granting permission for every inquiry. A person accused of something does not receive veto authority over investigation by withholding consent.
The instrument used by the investigation still has to be audited.
An unpublished manuscript crossed another boundary.
- Who supplied it?
- What authority did that person possess to transmit it?
- What happened to the manuscript inside the service?
- What did the resulting score actually establish?
- What error model travelled with it?
- What claims could responsibly be made from it?
- What weight did the recipients assign it?
Those questions do not disappear because the upload occurred in pursuit of truth.
The old article asked publishing to provide symmetrical provenance for its own machinery.
This case extends the rule.
The provenance auditor has provenance too.
An investigation into whether an author improperly introduced machine processing into a manuscript cannot treat its own machine processing of the manuscript as ethically invisible.
And this detector report did something much more important than "classify text."
This report travelled.
- A private suspicion is unstable.
- A report is portable.
- A report can be forwarded.
- Attached.
- Summarized.
- Shown to somebody else.
- Remembered as a "finding" after everybody has forgotten which assumptions produced it.
- A report can be forwarded.
Now, the classification has causal reach. That score does not sit beside the manuscript. It enters the manuscript’s future.
The first article said suspicion had become active information.
Falade’s account supplies us the missing machinery.
The Keystroke Alibi Has Arrived.
The first Modal Path Ethics article also made one extremely specific warning:
A writer should not need a keystroke alibi.
Jerry Falade has now independently arrived at the nightmare.
He tells writers to keep their old files.
- Keep the drafts.
- Keep the records.
- Keep the evidence.
Then he jokes that the strongest possible defense would be livestreaming the entire act of writing.
That joke works because the incentive is already fully real.
The publishing industry completely failed to build a provenance instrument. So now provenance becomes private survival labor.
The writer becomes their archivist.
The writer becomes their records officer.
The writer becomes their forensic laboratory responsible for preserving proof that the writer existed while writing.
This burden will not distribute evenly. Publishing already fully knows that.
A famous author already possesses provenance by reputation.
There are previous books. Interviews. Editors. Correspondence.
Years of public language against which a disputed manuscript can be compared.
A debut writer has less accumulated institutional memory.
A multilingual debut writer whose rhetorical habits already sit farther from an evaluator’s expectations now begins even farther back.
Now, let's add process variation.
- One writer keeps twenty-seven drafts.
- Another just overwrites one document.
- One composes in Google Docs.
- Another writes longhand.
- One dictates.
- One drafts sections in messages to themselves at midnight.
- One deletes everything because some failed drafts contain things they never wanted retained.
None of these is a theory of authorship.
They are just ways people write.
A mature provenance procedure can accept representative evidence when a particular concern arises. It cannot convert the absence of workplace surveillance into evidence that the worker never existed.
That would solve publishing’s missing instrument by attaching a panopticon to every writer.
Very clean.
Very efficient.
Completely indefensible.
The Appeal Had to Leave the Building.
Falade’s essay ends somewhere the first article could not.
The path reopened.
He announced that David Vigliano is now his literary agent. He has also been invited to discuss the controversy at the Aké Arts and Book Festival.
That matters structurally even before we know what commercial future Call Me, I’ll Hide the Body ultimately receives.
- Correction existed.
- Outside the original chain.
Falade did not describe an internal provenance tribunal receiving the disputed evidence, applying a serious published standard written by an adult, distinguishing research use from expressive generation, and issuing a revised finding.
He found another route instead.
- Another editor listened.
- Another agent accepted representation.
- Another institution gave him a platform.
That is a form of plural correction.
It is also expensive.
- A writer whose accusation becomes famous may attract people willing to reopen the case.
- A writer whose accusation silently kills three submissions may simply disappear.
Falade makes exactly this point in his essay: high-profile cases become public controversies; unknown writers can lose opportunities without anybody ever learning their names.
Exit is therefore a useful safety valve.
It is a terrible substitute for procedure.
The person who has been wrongly classified should not need to become interesting enough for another institution to rescue.
The Ruling.
The first ruling survives.
Authorship is a path claim.
Falade has now given us his account of that path.
On that account, the expressive work was his.
Artificial intelligence entered as a research instrument.
Yorùbá rhetorical traditions entered as part of the voice.
Suspicion entered later.
Detection entered later.
The detector report entered later.
The collapse entered later.
Then outside correction entered later still.
The sequence matters. Calling all of that AI use destroys precisely the information an "authorship inquiry" exists to recover.
There is also a new ruling.
A machine trained to imitate human language cannot become the court that decides which humans sound human enough.
Resemblance is not provenance.
Cultural unfamiliarity is not provenance.
A detector score is not provenance.
An em dash appears to have somehow survived centuries of human writing before ChatGPT. We can probably release it back into the wild.
Publishing still has a legitimate problem.
Human authorship matters. Disclosure matters. Rights matter. Generated prose matters. Candor matters.
Except none of those concerns require this industry to pretend that every new instrument capable of producing a fucking number has jurisdiction over the question.
The repair remains precisely what it was in August, with several patches now written in blood.
- Distinguish research from expressive generation.
- Require specific concerns instead of stylistic weather.
- Treat automated detection as evidence with limits rather than a verdict.
- Bring cultural and linguistic competence into any textual inquiry whose alarm is unfamiliar style.
- Keep evidence requests bounded to the disputed transition.
- Do not require total process surveillance.
- Audit how suspicion is distributed.
- Disclose what happens when an unpublished manuscript is transmitted into an external analysis system.
- Provide an appeal path before the writer has to leave the institution and reconstruct a career somewhere else.
Most importantly:
Do not make human beings deform their language to remain legible as human beings you absolute fucking imbeciles.
That is where the detector stops protecting authorship and starts rewriting it.
The machine learned from us.
If we then amputate every human habit the machine successfully learned to imitate, we are letting the imitation take possession of the source.
Jerry Falade should not have to become less Yorùbá so that a classifier can become more confident.
No writer should have to become worse evidence of themselves for a fucking auctioneer classifying art through a machine.
- The copy accused the source.
- The burden of proof belongs on the instrument.