The ethics of AI-assisted writing: disclosure, authorship, and what readers deserve

Why a general 'is AI ethical?' question fails

Asking whether using AI for writing is ethical, in the abstract, produces no useful answer because the question hides at least five different questions. Does the reader know the work was AI-assisted? Whose name belongs on the work? Are the factual claims true? Where did the underlying knowledge come from? Did the people whose writing trained the model consent? Each of these has a different shape and a different right answer. Treating them as one question is how the conversation stays stuck.

Question one: disclosure

Disclosure is the simplest of the questions and the most often dodged. The principle: readers are entitled to know, in non-buried form, whether the content they are reading was AI-assisted and what that assistance involved. Disclosure is not 'this article was written by AI' as a confession; it is 'AI was used for [specific stage] and the work was [reviewed / edited / fact-checked] by [named human].' Specific disclosure builds trust. Vague boilerplate ('this site uses AI tools') is closer to no disclosure. The right disclosure for an article that was AI-drafted and human-edited is different from the right disclosure for a piece where AI was used only for grammar checks. The right disclosure for a marketing landing page is different from the right disclosure for a news story. The general rule: disclose at the level of detail a reasonable reader would want before deciding how much to trust the piece.

Question two: authorship

Authorship is harder. The convention that 'the author is the human who wrote it' is doing a lot of work, and AI strains it. Practical positions in current use: authorship belongs to the person who originated and is accountable for the work, regardless of which tools were involved (the dominant editorial position); authorship requires substantive human creative contribution at the writing stage (the position taken by many academic journals); AI cannot be an author because it cannot bear accountability (a position with broad consensus); the model's training data contributors are not authors of any specific output (also broad consensus, though contested). The coherent position is that authorship is about accountability and substantive contribution, not about whether assistive tools were used. A writer who originated, edited, fact-checked, and stands behind a piece is its author even if the first draft came from an AI. A 'writer' who pasted a prompt, took the output, and shipped it without substantive engagement is closer to a publisher than an author — and the byline should reflect that honestly.

Question three: accuracy

The accuracy obligation does not change with AI use. If anything, it intensifies, because AI-drafted content fails on accuracy in different ways than human-drafted content. The writer who publishes a hallucinated citation is responsible for that citation, not the model. There is no diminished standard for AI-assisted work and there should not be. The practical implication is that the verification pass described in the long-form editing guide is not optional — it is the minimum standard for ethical publication.

Question four: attribution

AI models are trained on enormous amounts of writing whose authors did not specifically consent to their use as training data, and who do not see compensation when the model's outputs draw on their patterns. This is a real ethical issue, not a manufactured one. The current legal landscape is unsettled and the right policy response is contested. What writers can do individually is limited but not nothing: prefer tools whose training practices are transparent; cite primary sources explicitly when drawing on others' work, even when the AI did not surface the citation; avoid using AI to closely paraphrase identifiable work in ways that would be plagiarism if done by a human.

Question five: consent and use

Some uses of AI-assisted writing affect parties who did not consent — generating 'first-person' content for someone else's byline, drafting outreach designed to seem hand-written, fabricating endorsements or testimonials. These are not new ethical problems; they are old problems (ghostwriting, deception, fraud) with AI as the new vector. The general standard: do not use AI to misrepresent the relationship between the writer and the audience. A ghostwritten executive blog post is conventional. A 'personal' message from a CEO that the CEO never saw is deceptive. AI does not change the line; it just lowers the cost of crossing it.

What about academic writing?

Academic ethics around AI are evolving quickly and are field-specific. The defensible defaults: disclose AI use in any submission where the institution or journal asks; do not use AI to generate text passed off as original analysis; treat AI assistance with proofreading, summarisation of one's own notes, or non-substantive editing the same way one would treat any editing tool; never use AI in ways the institution's policy explicitly prohibits, even when you believe the prohibition is poorly drafted.

What about marketing and persuasive writing?

The ethical question for marketing is not 'is AI use disclosed?' (most readers correctly assume marketing copy is professionally produced) but 'are the claims honest and the relationship transparent?' A landing page partly drafted by AI is not deceptive. A landing page that misrepresents product capabilities is — whether the misrepresentation came from a human writer or an AI prompt. The ethical bar for marketing is unchanged by AI; the temptation to lower the bar because content is cheap is what to resist.

What about news and journalism?

News carries the highest disclosure obligation. Readers of news content reasonably assume named human reporting unless told otherwise. Any AI use in news writing should be explicitly disclosed and confined to roles that do not undercut that assumption — summarisation, translation, transcription, structured data tasks. Generated content marketed as reporting is the failure mode and the reason several major outlets have published explicit policies.

A simple framework

For any piece of AI-assisted writing, before publication, answer five questions explicitly: Will the reader know AI was involved at the level of detail they would want? Whose name belongs on this and on what grounds? Have I verified every factual claim? Have I credited any specific sources the work draws on? Does the piece misrepresent the relationship between the author and the audience? If all five answers are clean, the work is ethically publishable. If any are not, the issue is identifiable and fixable.

The harder cases

Some cases are genuinely hard. Ghostwriting with AI assistance for someone else's byline — where the named author has reviewed but did not write — sits in a long tradition with new tooling; the ethical line is reader expectation in that genre. Translation and accessibility uses of AI sit clearly on the ethical side because the alternative is exclusion. AI-generated images alongside human-written text raise their own questions about consent and attribution that this article does not attempt to resolve. The framework above does not eliminate the hard cases; it helps separate them from the easy ones.

Where this is heading

The professional norms around AI-assisted writing are forming now. Publications that lead with clear, specific disclosure and high editorial standards are setting the bar; publications that hide AI use or skip verification are accumulating risk. Over a five-year horizon the gap between the two will be more visible to readers than it is today. Writers and publishers who treat the ethical questions as load-bearing rather than ornamental will end up with the credibility — and the audiences — to match.

Frequently asked