What Is the Best Practice for AI Disclosure by Authors?
The best practice in September 2026 is to disclose AI assistance accurately, proportionately, and in a form the reader, editor, agent, or reviewer can understand. State which tools you used, what they did, and whether you independently verified or substantially rewrote the affected text. Do not simply paste a generic sentence claiming that “AI may have been used,” because that can be as misleading as saying nothing. A useful disclosure identifies the human author’s responsibility for the final work, while recognizing that different AI systems require different levels of reporting.
Also worth reading: What AI contract clauses do authors need to protect their manuscripts from machine learning training and unauthorized generation? · What is the C2PA implementation guide for authors, and how do writers add Content Credentials to their books and images? · What are backlist AI licensing rights and how can authors and publishers license older books for AI training?
There is not yet one universal disclosure form covering every book, article, or creative work. Oxford University Press has updated guidance for academic authors, Amazon has introduced policy concerning AI-generated book content, and newer proposals attempt to organize disclosures into increasingly specific levels. The Show HN description of ATS-1.0 calls it a six-tier technical standard for AI authorship disclosure, while other publishing initiatives use bylines, AI labels, or statements supplied at submission. These approaches share a basic principle: the more AI contributed to expressive or factual content, the clearer the disclosure should be.
For an author, the safest approach is to disclose known material AI use rather than wait for a rule to define exactly which prompts are exempt. This does not mean every spelling correction or brainstorming conversation must become a confession. It means that generated passages, translated prose, summaries, illustrations, research retrieval, and substantial structural edits should be documented internally at minimum, and disclosed externally when a publisher, journal, contest, or platform asks. The central question is not whether AI was present in the room, but whether its influence on the finished work would matter to an informed reader.
A defensible short disclosure might say: “Generative AI tools were used to suggest alternative phrasing and identify structural gaps. The author reviewed all suggestions, verified factual claims against primary sources, and independently rewrote the final text.” For heavier assistance, specify the function: “AI tools generated portions of the initial draft from the author’s outlines. The author edited, fact-checked, and reorganized all output and remains accountable for the published text.” Precision is more useful than a technical inventory of model names, subscription tiers, or prompt counts.
Why Publishing and Research Policies Changed
Publishing policy moved from a largely voluntary norm toward explicit reporting because readers, reviewers, and commercial partners could not reliably infer how text and images were produced. The research context supplied for this article describes Oxford University Press finding that AI use is “measurably” under-disclosed. The Bookseller likewise reported on updates aimed at making disclosure more accurate, while Research Information covered the revised Oxford guidance. These reports point to a reporting problem rather than proof that every undisclosed use represents deliberate deception.
The concern is accountability. A sentence drafted by a model may contain invented references, misplaced quotations, or confident errors. A disclosure does not remove those defects, but it tells an editor that another kind of review may be appropriate. It also helps a reader interpret voice, translation, imagery, and stylistic choices. Academic publication has developed authorship conventions because contribution matters; AI-assistance reporting extends that logic by describing the origin and extent of machine contributions.
The policy picture remains fragmented. CDC guidance on disclosing generative AI use in scientific work, Oxford’s author guidance, Amazon’s content requirements, and proposed standards such as ATS-1.0 address different settings and audiences. A technical six-tier proposal may be useful for institutions that need a common taxonomy, but it is not automatically a law or a substitute for a publisher’s instructions. A blanket requirement applied to every author would also be difficult to enforce, particularly after editing, source files, and version histories have passed among collaborators.
Outside scholarly publishing, the July 2026 OpenAI–Hugging Face security incident described in the provided Reuters research context reportedly went unnoticed for a week. The lesson for authors is narrower than a general claim that AI security is defective: an event can remain outside a publisher’s or platform’s knowledge until someone reports it. Authors should therefore maintain records of material AI use and be prepared to disclose it during submission, contract review, or a later inquiry. Policy is strongest when transparency creates an early record rather than an admission discovered after publication.
What Authors Need to Disclose—and What They Do Not
Start by separating assistance from authorship. Brainstorming titles, asking a general question about punctuation, or using a grammar tool is not equivalent to supplying a model with a chapter outline and publishing its generated draft with minimal revision. The difference lies in control, contribution, and the proportion of final material affected. A disclosure should communicate those distinctions without pretending that authorship can always be measured by a clean percentage.
Authors should record four practical elements: the tool or service, the task assigned, the stage of creation, and the human checks performed. “ChatGPT helped with research in March 2026” is weak because it does not say whether the tool summarized sources, proposed citations, or composed prose. “The model summarized author-provided source notes; the author checked every quotation against the original document and rewrote the passages” is much stronger. Dates and file versions matter when an editor asks how AI was used during a particular revision.
Material disclosures ordinarily cover generated or substantially rewritten prose, machine-assisted translation, synthetic images, altered author voices, AI-produced summaries or cover copy, and factual claims generated without reliable verification. Disclosure may also be required for research assistance if the tool selected sources, synthesized findings, or drafted a literature review. By contrast, routine spelling correction, accessibility tools that read text aloud, and personal notes that never entered the manuscript may not need public explanation. Authors should still preserve records if a contract defines those uses as reportable.
The proposed tiers should not be read as a moral ranking of authors. AI can be a useful aid for a visually impaired writer organizing notes or a researcher testing plain-language alternatives, while heavy generation can undermine a commissioned biography or academic article. Transparency should reveal what happened without turning an ordinary publishing tool into an accusation. If a rule appears overbroad, an author can request a written interpretation from the editor or rights-and-permissions team and note the agreed wording in the manuscript file.
A Practical Disclosure Workflow Before Submission
The first step is to classify the project by its destination. An academic article, self-published novel, children’s title, translated edition, and commercial nonfiction book may face different requirements. Read the submission guidelines, publishing agreement, journal policy, and platform terms in full. Check whether disclosure is requested in the manuscript, cover letter, metadata, image file, contributor statement, or a dedicated portal. A disclosure hidden in a file that editors do not open may technically exist but fail its communicative purpose.
The second step is to reconstruct the workflow from records rather than memory. Keep a dated document listing each material tool, its purpose, affected chapters, and the author’s editing or fact-checking actions. Retain prompts, outputs, reference notes, and relevant version histories where contracts or institutional policy permit. Do not upload confidential manuscripts, unpublished client material, personal data, or copyrighted source collections to a public service merely to complete a disclosure exercise.
The third step is to draft a plain-language statement before delivery. Identify the author, name the class of tools used, describe the tasks, and state the review performed. Avoid absolutes such as “the book contains no AI” unless the author can genuinely support that claim across drafting, editing, translation, artwork, and metadata. For lightly assisted work, a concise methods note may suffice; for generated drafts or synthetic visuals, use a dedicated disclosure tied to the affected component.
Finally, obtain human review appropriate to the risk. A developmental editor can check structure, a copyeditor can examine language, and a subject expert can verify claims. If AI altered quotations, citations, numbers, names, or descriptions of real events, compare them with the underlying sources rather than trusting fluent output. Update the disclosure whenever the manuscript is substantially regenerated. The record should show the final state of the work, not imply that later human editing automatically erased an earlier machine contribution.