What Counts as AI Assisted Writing for Disclosure Purposes

A writer runs a finished paragraph through a grammar checker and fixes three comma splices it flags. A second writer stares at a blank page, asks an LLM to write the paragraph outright, then lightly edits what comes back. Both writers used AI. Only one of them has something to disclose. That distinction, not the mere fact of AI involvement, is what every major publishing policy now turns on: if a human wrote the text and AI refined it, disclosure is generally not required; if AI produced the text, disclosure is required no matter how heavily the human later revised it. Policies from academic journals, Amazon's self-publishing platform, newsroom standards, and federal advertising regulators read differently on the surface, largely because each writes its own vocabulary and lists its own examples. Underneath, they are running the same test on text origin. Once a writer sees that test, the specific rule at any given publisher stops looking like an arbitrary item to memorize, because it is just one more application of a single, stable principle.

Assistive AI and disclosure requirements

Assistive AI works on text a person has already written. It corrects spelling, tightens punctuation, suggests a cleaner sentence structure, or reformats a document, without adding ideas, claims, or sentences that did not originate with the author. Publishers treat this kind of use the way they'd treat a spell-checker or a style guide: a tool that sharpens expression, not one that supplies it. A policy captured during the 2026 publisher rewrite cycle, drawn from a small newspaper's internal guidelines, calls this the calculator test: nobody labels the use of a calculator, so nobody needs to label mechanical cleanup like spell-checking either. That framing has become close to a consensus position across the industry. A major self-publishing platform applies the identical logic to books: help with ideas, grammar, or outlines, where the author wrote the actual text, counts as AI-assisted content and does not need to be disclosed to the platform. Only AI-generated content triggers that requirement. For most writers, this should be reassuring. You can run a draft through Grammarly's basic correction features, use a word processor's built-in spell-checker, or ask an editor to tighten phrasing, and none of it puts you on the disclosure side of the line, because none of it changes who actually wrote the words.

Generative AI and disclosure requirements

Generative AI is the origin of the text it produces: when a writer asks an LLM to draft a paragraph, summarize a stack of sources, translate a manuscript, or generate a list of references, the tool, not the person who prompted it, is the origin of that content. That shift in origin, not the sophistication of the tool or how much editing follows, is what triggers a disclosure obligation. The range of tasks that count as generative is broader than many writers assume. It includes drafting or rewriting sections of a manuscript, summarizing literature, generating or suggesting citations, creating figures or data visualizations, writing or debugging code used in data analysis, translating substantial portions of a manuscript, assisting with statistical analysis, and even drafting cover letters or responses to peer reviewers. Chigwada and Ngulube (2026) confirm that Elsevier requires disclosure for any use of generative AI or AI-assisted tools, a requirement that extends past the prose itself into figures and images, and that Springer Nature goes a step further by prohibiting generative AI images in its publications. Translation deserves particular attention, because it catches authors off guard more than almost any other case: under Amazon KDP's definitions, a translated manuscript counts as generated content even when every idea in it is entirely the author's own, since every word of the translated text was produced by the tool. The ICMJE updated its recommendations in April 2025 to require explicit disclosure of AI use in manuscripts, and Chigwada and Ngulube (2026) note that COPE and WAME hold similar positions independently, alongside it.

Academic Publishers' Submission Rules for Disclosure

The clearest shift in academic publishing policy during 2026 is not that disclosure is required (that was already largely true) but that policies now specify where the disclosure has to go. First-generation AI policies just told authors to disclose; the current generation tells them in which section, in what form, and with what supporting statement. Cell Press and Elsevier illustrate this well. Their submission rules require disclosure inside a dedicated declaration, filed alongside competing-interest and funding statements, and that declaration has to name the specific tool, explain why it was used, and confirm that the authors reviewed and edited the resulting content and take full responsibility for it. Elsevier adds two further constraints: AI-generated text must be reviewed and edited by the authors before submission, and AI use during peer review is restricted to narrow, assistive tasks, such as improving the language of a review report the reviewer has already written, using only private AI tools, with manuscript content never shared with any AI system. CASRAI (2026) and Chigwada and Ngulube (2026) both confirm this set of requirements. Cell Press editors may also request that authors retain their prompts and the AI outputs they generated, and Nature recommends the same practice, which functions in effect as a multi-year retention obligation. None of this is exhaustive of every academic publisher's rules, and it shouldn't be treated as a substitute for reading the actual submission guidelines of a target journal. Publisher-level policy sets a floor, but a journal's Instructions for Authors can impose stricter requirements on top of it, so check both before you submit.

Disclosure rules outside academic publishing: trade books, journalism, and commercial content

The same origin test that governs academic submissions appears independently in trade publishing, newsrooms, and advertising regulation, which suggests it reflects something structural about how writing and responsibility relate to each other rather than a convention specific to scholarly journals. On Amazon's KDP platform, the rule tracks the academic version closely: AI-generated text, images, or translations require disclosure through a checkbox in the publishing form, while AI-assisted content, meaning work the author wrote with AI only helping refine it, does not. That KDP requirement has existed since September 2023, and enforcement around it has tightened noticeably through 2025 and 2026. The Authors Guild has built its own, more explicit standard around the same line with a trademarked "Human Authored" seal. Its guidelines permit AI-powered spelling and grammar tools and allow a de minimis role for generative AI, such as building an index, but they do not permit AI to generate text that appears in the finished book; mentions of brainstorming, outlining, and creating auxiliary book parts were removed from the guidelines in May 2026, narrowing what counts as acceptable assistance. The Big Five trade publishers, Penguin Random House, HarperCollins, Simon & Schuster, Hachette Book Group, and Macmillan, have issued no public, house-wide rule restricting how authors use AI tools while drafting; their visible AI policy work has gone into training and licensing protections instead. Three of the five have added contract language barring the use of an author's work to train AI models without consent, but that only protects the author's existing work; it does not say how a new manuscript gets written. In journalism, one major news organization's July 2026 newsroom standards permit AI for headline ideation, document summarization, transcription, translation, grammar, and search optimization, while holding reporting, sourcing, editorial judgment, and verification as strictly human responsibilities, with a journalist reviewing and editing every AI output before publication. A federal regulatory agency applies a version of the same origin test to advertising: AI-written copy, AI-generated images, AI translations, and AI-created video all require disclosure, while grammar checking and analytics tools do not. In late 2025, the FTC issued its first warning letters under the Consumer Review Rule targeting fake reviews, including AI-generated ones, and at the same time vacated its earlier action against the AI review-generator Rytr, a move that reads as a shift toward evaluating how AI is used rather than treating AI tools themselves as inherently suspect. For a working creator, the practical shape of this rule is simple: writing a script, recording it in one's own voice, and using AI only for caption cleanup, color grading, or B-roll selection counts as AI-assisted production, and no disclosure is triggered.

Tool convergence, translation edge cases, and the Grammarly problem

The origin test holds up logically, but it has gotten harder to apply, because individual tools increasingly mix assistive and generative functions inside one interface. A policy that simply exempts "grammar checkers" no longer draws a clean line, since a tool marketed as a grammar checker may also offer a feature that drafts whole sentences or paragraphs from a short prompt. Grammarly is the clearest example: its basic spelling and grammar correction functions are assistive by any reasonable reading, but its Compose feature generates original text, which makes it generative under the same policies. A single product can sit on both sides of the line depending on which feature you actually used, so exemptions written at the tool level stop working, and you have to rewrite them at the feature level instead. Translation presents its own sharp edge. Most policies, including Amazon KDP's, classify translation as generative, because every word of the translated output was produced by the tool, regardless of how entirely human the underlying ideas and original-language text were. Many authors intuitively think of translation as refinement, a kind of polishing of work they already did, but the policies treat it as new text with a new origin. The scale of the compliance gap this ambiguity has produced is hard to ignore. A large-scale analysis cited in the 2026 publisher rewrite discussion found that only a tiny fraction of tens of thousands of papers published since 2023 explicitly disclosed AI use, even though most journals had adopted some form of AI policy by then. Either AI use in scholarly writing is vanishingly rare, which is not a credible reading of the data, or the first generation of binary disclosure policies was too vague for authors to apply consistently. Self-classification remains the weak point in the system: failing to disclose is hard to catch, over-disclosure carries no penalty, but an undisclosed AI contribution that later comes to light can be treated as research misconduct.

Structured, task-level disclosure in practice

The industry's answer to these blurred edges has not been to throw out the origin test but to make it operate at a finer grain. Instead of asking a single yes-or-no question, "did you use AI?", publishers are moving toward a structured question with four parts: which tool, for which purpose, in which section of the work, verified by whom? That structure typically asks a writer to name the purpose from a defined list, such as language polishing, translation, literature search, code generation, data extraction, image analysis, hypothesis generation, or draft writing; to define the scope, meaning which sections or artifacts the AI touched; to include a statement that a named human verified and takes responsibility for the result; and to retain the prompts and outputs used, in case an editor needs to review them later. A major wire service's July 2026 newsroom standards show the same shift from the journalism side: they enumerate specific tasks and assign each one its own rule instead of debating AI's acceptability as an abstract category. Summarizing a document a reporter has already obtained is permitted. Deciding what that document means and reporting on it is a human responsibility that AI cannot substitute for. This is the model that resolves the question of how to treat grammar-checking tools versus translation tools without abandoning the underlying logic: a feature-by-feature, task-by-task accounting replaces a single checkbox, and the origin test still decides each individual answer.

A working decision framework for writers navigating disclosure today

If you face a new publisher or platform, you do not need to learn an entirely new rulebook, because one underlying test applies consistently, and only the venue-specific details change on top of it. The starting move is to apply the origin test to every AI use in the workflow, asking two simple questions for each one: did a person write the text and AI only refine it, in which case it is assistive and disclosure is generally not required, or did AI produce the content even if a person edited it afterward, in which case it is generative and disclosure is required. That test has to be applied at the level of the specific feature used, not the tool as a whole, since a single product like Grammarly contains both an assistive spell-check and a generative Compose function governed by different rules. The same logic extends to visuals and code: a figure, a data visualization, or a block of analysis code each carries its own disclosure trigger and has to be assessed on its own, separately from the surrounding prose. Once that assessment is done, the next move is to check the actual requirements of the specific venue before submitting anything. Academic authors should check both the publisher's general AI policy and the target journal's Instructions for Authors, since journal-level rules can be stricter, and should confirm exactly where a disclosure needs to appear, whether that's the Methods section, the Acknowledgements, a dedicated declaration, a cover letter, the submission form, or more than one of these at once. Authors working in trade publishing should check whether the platform requires a private disclosure, the way KDP does through its checkbox, and whether a certification program like the Authors Guild's "Human Authored" seal applies to the project. Writers producing commercial content should apply the FTC's version of the same test, disclosing any AI-produced text, image, translation, or video used in advertising while leaving grammar-checking and analytics tools undisclosed. When a disclosure is required, the emerging standard asks for five elements: the name and version of the tool, the purpose it served, drawn from a recognized list such as language polishing or draft writing, the scope of what it touched, a statement that a named person verified the result and takes responsibility for it, and retention of the prompts and outputs for whatever period the venue specifies. This framework is worth learning rather than just looking up each time because it rests on the origin of the text, not on the name of a tool or the wording of a particular publisher's policy. Tools will keep changing, and individual publishers will keep rewriting their rules, just as the 2026 cycle has already shown. When that happens, a writer only needs to recheck where and how to disclose for that specific venue. The test for whether disclosure is owed in the first place does not change.