The Manuscript Was “Genius.” Then Suspected AI Use Made It Unpublishable

Editorial image generated by the author.

What the collapse of a multimillion-dollar publishing deal reveals about authorship, artistic value and the price placed on a recognisable human process.

A manuscript does not often lose millions of dollars in value without changing a word.

Jerry Falade’s debut crime novel, Call Me, I’ll Hide the Body, reportedly attracted a fourteen-way auction in the United States and an offer exceeding $2 million from Minotaur, an imprint of Macmillan. Several six-figure bids were also reported in Britain. Editors had read the novel, discussed it internally and decided that it was not merely publishable but unusually valuable.

Then Falade’s agents withdrew the manuscript. According to the agents’ explanation, an editor had raised concerns about possible artificial-intelligence involvement, and the agency concluded that it could no longer authenticate how the manuscript had developed from its origins to the version being offered to publishers. The agent who withdrew it continued to describe the book itself as “genius”.

Falade denies using AI to write the novel and has argued that the suspicion reflects racial bias within publishing. He later said that he had used a locally run model for research, including questions related to the novel’s criminal and forensic details, but not to compose the manuscript. Public reporting has not revealed enough evidence for outsiders to determine how the suspicion arose or whether it was justified.

The case should therefore not be described as the exposure of an AI-written novel. Its importance lies in the contradiction it created. The manuscript had already passed the quality test. What failed was confidence that the text, the named author and the process behind it belonged together.

A publisher does not buy only a book. It buys a claim about where the book came from, who can take responsibility for it and whether the person named on the cover can produce another.

What a Publisher Buys Beyond the Manuscript

A cultural work carries several kinds of value at once. It is an artefact: a sequence of words, sounds or images capable of producing pleasure, insight or emotional response. It is also an attributed act. The author’s name suggests that a particular person imagined the work, selected its language, revised its failures and decided when it was finished.

The work also belongs to a production system. That system determines who contributed, who was paid, what rights can be transferred and who remains responsible if the finished object contains copied, defamatory or otherwise unusable material. These forms of value usually arrive together. Generative AI pulls them apart.

A novel can be excellent while its authorship remains uncertain. It can contain substantial human judgement without every sentence having originated directly from the named writer. Someone can also type every word personally and produce something lifeless and derivative. The amount of manual labour involved cannot by itself establish either quality or authorship.

For a publisher, however, the author’s role has practical value beyond the completed manuscript. A major acquisition may involve extensive revisions, another novel in the same voice, public appearances, screen rights and years of cooperation with the person whose identity will help sell the book. Editors may demand a different ending, the removal of a subplot or a substantial shift in emphasis. They need confidence that the author can understand those requests and reproduce whatever made the original manuscript valuable.

A prototype proves that one object exists. It does not prove that the producer can manufacture another, modify the design or maintain the same standard under pressure. A book can therefore be excellent and still be commercially unreliable. The publisher is buying the author’s capacity to remain an author.

Publishing has never required every word to emerge from one unaided mind. Editors rewrite sentences. Book doctors repair structures. Co-authors divide labour. Ghostwriters turn interviews, notes and spoken recollections into books credited to someone else. These arrangements are commercially manageable when the contribution is understood, authorised and reflected in the contract. Difficulty begins when substantial delegation is concealed or cannot be reconstructed.

The relevant question is not whether outside assistance occurred. It is whether the contribution structure matches what publishers and readers have been told they are buying.

Assistance, Collaboration or Substitution?

The comparison between an AI system and a human editor is imperfect, but it cannot simply be dismissed. A competent editor identifies repetition, exposes structural weakness, questions a character’s motivation and notices when a scene begins too early or ends too late. Good editing helps a writer see a manuscript that has become too familiar to judge clearly. Far from disqualifying the author, editorial support is normally treated as part of serious publishing.

Such assistance is also expensive. Established writers and contracted authors can obtain professional scrutiny that many aspiring writers cannot afford. Others depend on friends, writing groups or whatever criticism they can persuade someone to provide. The romantic image of the solitary novelist conceals a material inequality: some writers can purchase better feedback than others.

An AI system can make a rough version of that service available almost immediately. It can look for continuity errors, compare possible structures, interrogate the logic of a chapter or produce alternatives to a weak sentence. Declaring every such use illegitimate would preserve unequal access to editorial help rather than preserve authorship.

Human editing and machine assistance nevertheless differ in important ways. An editor can explain a judgement, defend an unusual choice and learn which apparent imperfections belong to the writer’s voice. Editors carry professional reputations and can be named, contracted and held accountable. A model may flatten eccentric language towards statistical regularity, introduce mistakes without recognising them or process confidential material under conditions the writer barely understands.

The largest difference concerns expressive contribution. Asking a system where a chapter loses momentum resembles requesting editorial criticism. Asking it to write the chapter delegates part of the composition. Between those poles lie brainstorming, sentence alternatives, generated descriptions, rewritten paragraphs, structural proposals and complete scenes later altered by the user.

Visual art presents the same continuum. Photoshop allows an artist to alter light, remove objects, combine images and construct scenes that never existed. Generative tools go further by supplying details the user did not individually choose. Manually changing the colour of a wall involves a different degree of expressive control from requesting an entire ruined city and selecting one of several outputs.

The presence of software does not settle the question. The more useful test is how much expressive judgement remained with the person claiming authorship.

Under current US copyright policy, the Copyright Office takes a similar approach. Human-authored material can remain protected when AI has been used as a tool, and human selection, arrangement or modification may itself be copyrightable. Purely generated material is treated differently, while prompts alone do not ordinarily provide enough control over the resulting expression to establish authorship.

Copyright law cannot tell us what counts as art. It does expose how little the sentence “AI was involved” tells us about who made the expressive decisions.

Who Gains When Creation Becomes Cheaper?

One objection to generative AI is that it allows people to produce work they would otherwise lack the skill, time or patience to complete. That is true, and it is one of the technology’s strongest claims to legitimacy.

A person may have a compelling story but lack the command of prose needed to sustain a novel. Another may understand composition but be physically unable to paint. A non-native speaker may possess greater narrative imagination than linguistic fluency. A parent working full time may not have a decade available to master every technical skill relevant to a project. AI cannot guarantee good art, but it can help people without money, fluency, physical capacity or abundant free time complete work that would otherwise remain unrealised.

Creative technologies have repeatedly reduced the cost of execution. Photography made image-making possible without the manual technique of portrait painting. Desktop publishing removed much of the specialised labour once required to typeset and distribute a document. Each expansion produced valuable work alongside enormous quantities of material that nobody needed.

Lower barriers expand participation and spam through the same mechanism: more people can produce more material at lower cost. Cheap creation therefore brings a discovery problem alongside questions of permission, ownership and authorship. Without effective filters, worthwhile work may become harder rather than easier to find.

Nor does the removal of traditional gatekeepers guarantee independence. Writers who once relied on editors, studios or publishers may instead become dependent on model providers, subscription services and computing infrastructure they do not control. Those broader consequences belong to the broader dispute over AI, copyright and cultural production. The Falade case presents a narrower problem: what happens when the provenance attached to one valuable manuscript can no longer be trusted?

The economic fears of writers, illustrators, translators, actors and musicians are not difficult to understand. Systems are being built to perform tasks from which they earn their living. Creative workers do not, however, possess a permanent right to remain untouched by automation. Technology has displaced skilled work in many industries, and artistic identity cannot by itself turn economic disruption into an exception.

Creative work does have unusual commercial features. Books, paintings and music are frequently marketed through the identity and biography of their maker. Models may be trained on existing works without consent or compensation. Generated output can imitate the recognisable manner of a living artist. Buyers may believe that they are paying for human testimony or expression rather than merely a functional product.

Automation does not answer questions of licensing, attribution, bargaining power or compensation. Artists need no mystical exemption from economic change to deserve protection against deceptive attribution and unauthorised commercial appropriation.

The Human-Authorship Premium

Opposition to AI-generated art is often presented as a judgement about quality. A study published in Judgment and Decision Making provides unusually direct evidence that attribution can alter how people evaluate an otherwise unchanged story.

In its first experiment, 1,682 participants read a short story described as either human-written or AI-generated. The AI-produced stories received higher ratings for quality and absorption, yet stories labelled as human-written were evaluated more favourably than stories carrying an AI label. In two further experiments involving 905 participants, readers were no better than chance at identifying which stories came from humans and which came from ChatGPT.

The experiment was deliberately narrow. It used three human-written stories and three corresponding ChatGPT stories, all approximately 1,000 words long. The researchers acknowledge that novels may produce different results because they require sustained character development and structural control across a much larger work. The study does not demonstrate that AI fiction is generally superior, or that current systems can produce a successful novel unaided.

Its more limited finding is still important: readers’ evaluations changed when the label changed, even though the words did not.

We might call the difference a human-authorship premium. Readers may prefer the artefact produced by a machine while preferring to believe that a person produced it.

That response need not be reduced to hypocrisy or virtue signalling. People often value qualities that are not visible in the finished object. A handmade object may command a higher price than an indistinguishable industrial version because craftsmanship and provenance form part of what the buyer wants. A live recording can carry meaning partly because musicians performed together at a particular moment.

Books can carry similar value. Readers may believe that fiction offers contact with another mind: someone imagined these people, struggled with the language, discarded failed versions and decided that this particular story deserved to exist. The labour does not guarantee quality, but knowledge of the labour alters the encounter.

In some forms, human origin is inseparable from the work. A memoir matters because someone lived through the events. A love letter matters because a particular person chose to send it. Identical language produced without the relevant experience may preserve the text while changing the communicative act.

Beauty and provenance are different goods. Some readers will care mainly about the first. Others will pay for both.

The Detection Trap

Once provenance becomes valuable, publishers need some way to authenticate it. The obvious solution is AI detection, but a detector does not observe how a manuscript was written. It infers origin from statistical characteristics of the finished text.

A 2023 study of seven detectors demonstrated the danger. Of 91 TOEFL essays written by non-native English speakers, 89 were flagged by at least one detector and 18 were classified as AI-generated by all seven. The researchers also showed that relatively simple alterations could help generated text evade the same systems.

The study examined tools available in 2023 and does not prove that every newer detector behaves identically. It does show why a probability score should not be treated as a record of authorship, especially when a career or multimillion-dollar contract depends on the result.

Falade has argued that race influenced the suspicion directed at his manuscript. The public evidence does not allow that allegation to be confirmed or dismissed. An opaque authentication process nevertheless leaves room for unequal scrutiny. Writers working outside familiar cultural and linguistic patterns may be asked to prove themselves in ways that others are not.

Informal suspicion creates another problem. Ordinary stylistic habits—regular paragraph structures, polished transitions, symmetrical arguments or conspicuous punctuation—can acquire reputations as signs of machine writing even though they long predate generative AI.

Writers may begin preserving awkwardness because excessive polish appears incriminating. Non-native speakers may avoid clear, regular English. Editors may hesitate to improve a passage that could become “too smooth”. Generated prose, meanwhile, can be instructed to include fragments, digressions, inconsistencies and deliberate roughness.

Machines would learn to imitate human imperfection while humans performed imperfection to evade machines.

It is doubtful that any stable textual signature can survive changing models, extensive revision and deliberate evasion. A passage may begin as a human draft, pass through several machine-assisted transformations and then be rewritten sentence by sentence. Another may emerge from detailed human planning but contain large blocks of generated prose. The final text cannot reliably reconstruct that history.

Why “Made with AI” Says Almost Nothing

Mandatory labelling is often offered as a compromise: permit AI-assisted work but require disclosure. Falade’s own account demonstrates why a simple label would be nearly useless. He says that he used a model for research but not composition. Would that make the resulting book an AI-assisted novel?

A declaration that a work was “made with AI” might refer to spellchecking, title suggestions or questions about pacing. It might also describe a manuscript generated scene by scene and lightly revised by the person named as author. Those processes involve different kinds of contribution and should not carry the same label.

A more useful disclosure system would identify the role performed:

• Mechanical assistance: spelling, grammar, formatting, transcription and similar operations that do not materially supply expressive content.
• Editorial assistance: criticism, structural diagnosis, continuity checks, research organisation and suggestions that the writer evaluates and implements.
• Generative assistance: discrete passages, images or other expressive elements produced by AI and incorporated into a substantially human-directed work.
• Generative composition: significant portions of the initial expression produced by AI, even where the human later revises and rearranges them.
• Directed generation: work in which the person’s principal contribution consists of prompting, selecting and arranging machine-produced outputs.

No taxonomy will convert authorship into a clean percentage. Human collaboration has never been so precise. Editors rewrite. Translators reshape. Ghostwriters construct entire books from another person’s ideas and recollections. Disclosure should make the contribution structure intelligible rather than calculate creative purity.

Nor would disclosure require a foolproof detector. Contracts already depend on declarations that cannot be verified independently in every case. Authors warrant that their work is original, that quotations are lawful and that they control the rights they are selling. Evidence becomes important when a dispute arises.

For a multimillion-dollar acquisition, drafts, notes and version histories may reasonably form part of due diligence. Requiring exhaustive process records from every writer would be far more troubling. Creative work is often nonlinear, private and poorly documented. A system built around constant proof of humanity could become more intrusive than the technology it is meant to regulate.

The Words Stayed the Same

Generative AI weakens the connection between visible quality and the labour required to produce it. Competent prose, polished illustration and plausible music can increasingly be created without the years of training once implied by the result.

That change will produce oceans of disposable material. It will also produce books and images that would otherwise never have existed, including some that people will value. Preserving difficulty for its own sake would protect existing hierarchies as often as it protects artistic standards.

The harder task is deciding what remains scarce when competent expression becomes cheap. Distinctive experience remains scarce. Judgement and accountability remain scarce. So does confidence that the person named as author can explain, revise and reproduce the creative decisions attached to the work.

Falade’s manuscript did not become worse when its provenance became uncertain. What disappeared was the industry’s confidence that the text, the named author and the future books belonged to the same creative process.

That confidence is not decoration around a publishing deal. It is part of what the publisher is buying. AI assistance need not contaminate a work, but authorship cannot become a ceremonial name attached to an unknowable production system.

The words remained the same. The promise attached to them did not.