The AI Scandal Isn't Authors Using AI. It's Services Hiding It.

Published October 26, 2026 · 16 min read

The central trust problem in AI-assisted publishing is not that a tool was used. It is that an author, editor, illustrator, or service may represent the work one way while doing it another. Recent reporting from Australia makes that distinction concrete: authors said they paid for professional publishing help without being told that AI was used for illustrations. The practical response is not another unreliable detector. It is written disclosure, informed consent, clear deliverables, and a contract that says who may use AI, where, and for what.

For much of the past year, the familiar “AI in publishing” story has pointed in one direction. An author is accused of undisclosed AI use, a prize or publisher responds, and a career is placed on trial. An ABC News investigation published October 25, 2026, points the other way. It asks what happens when authors buy editing, illustration, or publishing services and the provider allegedly fails to explain how AI entered the work.

What ABC News reported

ABC News reported that several debut authors who paid Brisbane-based Patrick White Publishing, also styled PW Publishing House, described editing problems, additional package offers after the initial quote, and illustrations they understood to be custom work. ABC said some covers promoted by the company returned positive results for OpenAI or Google Gemini watermarks. It also said the author agreements it reviewed did not mention AI-generated illustrations.

The report included other serious allegations. One author said she was asked to remove a negative Facebook post as part of a refund discussion. ABC said it saw a separate email in which the company asked another author to remove a negative review before work would continue. ABC also reported that the Australian Business Number displayed on the company's materials belonged to an unrelated business, and that storefront images attached to listed addresses were AI-generated rather than photographs of the premises.

Those claims require careful language. They are allegations reported by ABC, not findings by a court. Patrick White Publishing did not respond to ABC's requests for comment. In a Trustpilot response quoted by ABC, the company said it disagreed with several claims and remained committed to transparency and professional communication.

ABC also interviewed a customer who said she was satisfied after paying for a broader package that included editing, marketing, print, ebook, and audiobook work. She did not object when ABC's check indicated that her cover illustration was AI-generated. That evidence matters: it shows why the issue cannot honestly be reduced to “AI art is always bad.” One customer may knowingly accept it; another may believe she is paying for a human illustrator. Consent changes the transaction.

The company's name can also invite confusion. ABC reported that it has no connection to Patrick White, Australia's Nobel Prize–winning novelist, who died in 1990. The literary executor of White's estate told ABC that she had asked the company to stop using the name.

The same trust problem can run in both directions

Set the Australian reporting beside the recent case of novelist Thélyson Orélien. On October 21, 2026, Éditions du Boréal said it was ending its relationship with Orélien after an external analysis of working documents found what the publisher described as AI markers in metadata. The publisher also said it still did not know the extent to which AI had been used in the novel. Orélien had previously denied using AI to write it while acknowledging limited use for research and translation. Several literary juries removed the book from consideration, while other prize organizations took different positions.

The two stories have different facts, evidence, and consequences. They should not be collapsed into a verdict about any person or company. Their common thread is narrower and more useful: a party trusted to describe a creative process is accused of describing it inaccurately or incompletely.

That is why “Was AI used?” is the wrong first question. The better questions are: What was promised? What was prohibited? What was disclosed? Who consented? What evidence exists? What remedy did the agreement provide if the representation was wrong?

Publishing policy is moving toward disclosure, not detection

Three developments in October 2026 point in the same direction.

• Prix Pascal de Littérature IA opened a prize for AI-co-written books. Its official submission requirements include a technical dossier identifying the model, agents, or prompts used. AI use is not hidden evidence there; disclosure is part of entry.
• Coach House Books told The Canadian Press that its contracts ask authors to affirm that submitted books do not contain AI-generated material. Its editorial director also said available detection tools have proved unreliable and problematic.
• The Authors Guild published publisher best practices on October 13, 2026, calling for written AI policies, express author permission before inserting AI-generated content or artwork, transparency about detector use, and no reliance on detectors alone.

These approaches are not identical. A prize can welcome disclosed co-writing, a publisher can prohibit generated prose, and an author can permit assisted proofreading but reject generated cover art. What they share is governance before accusation: define the boundary, obtain agreement, and preserve evidence.

What an AI disclosure should actually cover

“We use AI sometimes” is not a useful policy. A service should connect the tool to the deliverable and the author should be able to accept or reject each use separately.

An edge case makes the need for precision obvious. Suppose a human illustrator draws a cover, then uses generative fill to extend a background. Is that a custom human illustration? It may be, but the author still needs enough information to make the platform disclosure and rights decision. “Custom” describes exclusivity or commissioning; it does not describe production method.

A practical pre-payment check for authors

1. Identify the legal party. Match the business name, registration number, invoice entity, payment recipient, and contract signatory. Australian authors can check an ABN through the government's ABN Lookup; use the equivalent official register in your jurisdiction.
2. Define every deliverable. “Professional editing” should say whether it means developmental editing, line editing, copyediting, proofreading, or a combination. “Custom cover” should specify the creator, method, revisions, source files, and licence.
3. Get a stage-by-stage AI policy. Ask about text, images, audio, translation, formatting, metadata, and marketing. A provider should not treat consent to one use as permission for every use.
4. Write down the price boundary. Separate optional add-ons from work required to receive the promised result. If a package expands after payment, pause before approving another charge.
5. Preserve versions and approvals. Keep the quote, agreement, receipts, manuscript versions, art proofs, revision requests, and emails. Do not rely on a portal as the only copy.
6. Control platform disclosure. If you will publish through Amazon, review the current KDP AI disclosure checklist against the final files—not the provider's marketing language.
7. Protect review independence. Read refund and dispute terms before paying. If a disagreement develops, keep communications factual and preserve evidence. Government consumer-protection agencies such as Australia's Scamwatch explain reporting options.

Turn the disclosure into an acceptance test

A policy is useful only if it changes what the author receives and what the author may reject. Before work starts, convert each promise into an acceptance condition. If the package includes human line editing, the condition might require tracked changes from a named editor, a short editorial memo, and one revision round. If the package includes original illustration, the condition might require the artist's name, a method statement, a commercial-use licence, and editable source files. If AI-generated art is allowed, the condition should say so directly and identify who is responsible for platform disclosure and rights review.

This matters because a polished final PDF cannot prove how it was made. A clean cover may come from a commissioned illustrator, a licensed stock composition, a generated image, or a mixture of all three. A well-written paragraph may be the work of an editor, a model, or both. The acceptance test should therefore examine the evidence trail, not merely whether the deliverable looks professional at first glance.

Consider a cover package sold as “custom illustrated.” That phrase is too elastic to accept on its own. One provider may mean an illustrator starts from a blank canvas. Another may mean a designer combines licensed stock assets. A third may generate the central image and customize the typography. The author does not need a moral slogan; the author needs the production method, rights chain, files, and disclosure obligation. The acceptance test fails if those facts arrive only after final payment.

The same logic applies to editing. A provider may legitimately use software to flag repeated words, compare a style sheet, or identify missing quotation marks while a human editor makes every editorial decision. That is different from sending the manuscript through a model and returning unreviewed rewrites. The contract should distinguish analysis from generation, and assistance from substitution. “AI-assisted editing” without that distinction leaves both parties guessing.

Failure modes that a one-line AI policy misses

The subcontractor gap. A service may promise that “we do not use AI” while outsourcing illustration, narration, translation, or marketing copy to someone operating under different rules. The agreement should bind subcontractors to the same disclosure and confidentiality terms and make the primary provider responsible for collecting their evidence. Otherwise, the author has a promise from one entity and a deliverable made by another.

The revision-round switch. The first draft may be created as promised, but urgent corrections can be routed through a generative tool to save time. That can change character details, introduce factual errors, or alter an image's provenance after the author approved the original method. Ask whether the AI policy applies to revisions, rush work, and substitute assets—not only the first version.

The training and retention ambiguity. A provider can avoid generating text and still upload an unpublished manuscript to an external system for analysis, transcription, translation, or quality checks. Authors handling memoir, confidential client material, embargoed research, or licensed characters need to know where files go, how long they remain, whether humans can review them, and whether they may be used to improve a model. A promise about outputs does not answer a privacy question about inputs.

The rights-chain mismatch. A contract may grant broad commercial rights even when the provider cannot document the source of an image, font, voice, or passage. Contract language cannot manufacture missing provenance. Ask for the underlying licence and creator record, and reserve the right to replace material whose source cannot be substantiated. This is particularly important when a distributor, retailer, insurer, or foreign publisher later asks for evidence.

The platform-answer handoff. Some service packages upload files on the author's behalf. That convenience becomes a risk if nobody agrees who answers AI-disclosure questions or preserves a copy of the submission. The author should approve the answers before upload and keep screenshots or exported records with the final manuscript package. If the platform changes its policy, the retained record shows what was represented at the time.

A realistic evidence workflow from quote to publication

At the quote stage, create a one-page deliverables schedule. List editing, cover, interior art, formatting, metadata, marketing, audio, translation, and upload support separately. Beside each item, record the production method, responsible person, permitted AI use, expected source files, approval point, and price. If a sales call contains an important promise, ask for it to be added to the schedule before paying.

During production, approve work in stages rather than as one final bundle. Review the developmental edit before line editing begins. Approve the cover concept before final rendering. Check a sample chapter after formatting. For each approval, save the file and date locally. This makes disagreements narrower: the parties can identify the stage that changed instead of arguing about an entire package.

Before final payment, run a delivery audit. Open every file rather than trusting a download list. Confirm that editable files are actually editable, fonts and images have usable licences, tracked changes identify the editor, and final metadata matches the book. Compare the evidence package with the disclosure answers required by the intended retailer or contest. If something is missing, request the record while the provider still has the project open.

After publication, retain a compact archive: signed agreement, invoices, approvals, final source files, licences, AI-use statement, platform submission record, and a checksum or dated copy of the published files. This is not an invitation to surveil every creative decision. It is ordinary chain-of-custody discipline for a commercial product. It helps a provider demonstrate honest work and helps an author correct a problem without relying on memory or detector scores.

There are also legitimate cases where the evidence will be incomplete. A sole editor may not have enterprise-style audit logs. A traditional artist may provide layered artwork but no time-lapse recording. An older stock licence may use broad language. The decision is proportional: identify which uncertainty affects disclosure, privacy, exclusivity, or commercial rights, and decide whether a written representation is enough. The goal is not perfect forensic proof. It is enough reliable evidence for the actual risk.

A practical decision rule is to separate reversible uncertainty from irreversible exposure. A typo found after delivery is inexpensive to correct. An ambiguous instruction about a brainstorming tool may be clarified in an email. By contrast, uploading a confidential memoir to a system with unknown retention, printing thousands of books with undocumented cover rights, or submitting inaccurate platform disclosures creates consequences that cannot be cleanly undone. Require stronger evidence before those irreversible steps. For low-risk assistance, a short written method statement may be enough. For generated artwork, licensed voices, sensitive manuscripts, or distribution performed in the author's name, require the source files, rights documents, approval record, and final disclosure answers before acceptance. If a provider cannot supply them, the author can narrow the scope, substitute a deliverable, delay publication, or choose another service. That is not hostility to AI; it is ordinary risk allocation. The provider knows what process it used, while the author carries the public credit, platform account, and long-term relationship with readers. A fair contract places the evidence duty with the party best able to produce it and gives the author a meaningful choice before the risky step occurs.

What transparent AI use looks like in practice

Consider a hypothetical children's author buying editing, cover, and formatting services. A transparent provider might say: grammar suggestions may use an automated tool, but a named editor reviews every accepted change; the cover is generated with an image model and then typeset by a designer; the author receives the final image, layered source file, and commercial-use terms; the manuscript is not used for model training; and the author approves any AI-generated material before the files are delivered.

The author may accept that package, reject it, or negotiate a human-illustrated cover. All three outcomes are healthier than learning the production method from a watermark check after payment. Disclosure does not force agreement. It makes agreement possible.

The same discipline applies when authors use AI themselves. Preserve your outline, prompts, notes, accepted drafts, and human revisions. Follow the rules of your publisher, contest, or platform. If you are using a long-form tool, build review points into the workflow rather than treating generation as the final manuscript. ShakespeareAI's guide to human review checkpoints and the AI-novel editing workflow show how to separate generation from editorial acceptance.

Where ShakespeareAI stands

ShakespeareAI helps authors develop their own books using AI for planning and generation. The author supplies the premise and creative direction, reviews the outline and chapters, decides what to keep, and remains responsible for editing and publication choices. The AI role is part of the product's identity, not a hidden substitute sold as something else. Current capabilities and limits should still be checked on the features page before starting a sensitive or contracted manuscript.

That commercial interest is exactly why concealed AI use harms the whole field. When a service sells generated work as human work, it taxes the credibility of authors who disclose their process. When an author is judged solely by a detector score instead of a clear agreement and reliable evidence, the industry creates a different kind of unfairness.

The standard should be simple

Do not turn every use of AI into a scandal. Ask whether the use contradicted a promise, policy, contract, or informed choice. For authors and service providers alike, the durable standard is the same: say what the tool did, obtain consent where another person's work or money is involved, preserve enough evidence to resolve a dispute, and correct the record when a description was wrong.

Detection software can raise a question. It cannot supply the missing agreement. Disclosure can.

Reporting concerning Patrick White Publishing is attributed to ABC News and includes allegations that remain contested in part. The Éditions du Boréal description is attributed to the publisher's October 21, 2026 statement. Sources were verified October 26, 2026.