When and How to Disclose That You Used AI

The question comes up constantly and gets answered badly in both directions. One camp says disclose everything, which collapses under its own weight — nobody discloses spellcheck. The other says disclose nothing, which is straightforwardly dishonest in some contexts and career-ending in a few.

The useful answer is neither. Disclosure is situational, and the situation is decidable if you ask the right question.

The question that decides it

Not "did I use AI?" That question produces absurd results, because the category runs from autocorrect to generating an entire document, and treating those the same is useless.

The better question: would the person receiving this work make a different decision if they knew?

That reframes disclosure around the audience's actual interest rather than around the technology. It also explains the obvious cases immediately.

A reader of a blog post about database performance cares whether the benchmarks are real. Whether the transitions were drafted by a model does not change how they should read the numbers. No disclosure needed.

A professor assessing whether you can construct an argument is measuring a capability. If a model constructed the argument, the assessment measures nothing. Disclosure — or rather, compliance with whatever the policy requires — is essential.

A client paying for original photography is buying provenance. A generated image, however good, is not the thing they purchased. Disclosure is mandatory.

Same tool, three different answers, and the question sorts them without any hand-wringing about the technology itself.

A rough scale of assistance

It helps to be concrete about what "used AI" means, because the disclosure obligation scales with how much of the thinking was delegated.

Mechanical assistance. Spellcheck, grammar correction, autocomplete of a phrase you had already decided on. Universally accepted, never disclosed. It has been part of writing for thirty years.

Language assistance. Rephrasing a sentence you wrote, tightening a paragraph, fixing register. Your ideas, your structure, your evidence — the machine adjusted the expression. This is the tier where most people's actual use sits, and in most contexts it needs no disclosure. It is also the tier where non-native English speakers rely on tools heavily and reasonably, and treating it as suspect creates a fairness problem covered in why AI detectors fail non-native English speakers.

Structural assistance. Asking for an outline, having a model organise material you supplied, generating section headings. The judgement about what to say is still yours; the arrangement was assisted. Grey area, and where context starts to matter a lot.

Substantive generation. The model produced the argument, the examples, or the evidence, and you edited. This is the tier where the work is genuinely collaborative, and where disclosure is usually appropriate outside of routine commercial content.

Wholesale generation. Prompt in, output out, minimal review. Disclosure is the least of the problems here, because nobody has verified the claims.

Most disputes happen because two people are talking about different tiers while using the same phrase.

Where disclosure is not optional

Some contexts have rules, and the rules are not always where people expect.

Academic work. Every institution now has a policy and they vary enormously — some ban AI entirely, some permit language assistance, some permit substantive use with citation. Read yours rather than assuming. Undisclosed use where the policy requires disclosure is misconduct regardless of how much the tool contributed. Our fuller treatment is in AI writing tools and academic integrity.

Journalism. Most newsrooms have adopted standards requiring disclosure of AI involvement in reporting, and essentially all of them prohibit generated imagery presented as documentary photography. The reporting-versus-drafting line matters here more than elsewhere.

Regulated professional advice. Legal filings, medical documentation, financial advice, engineering sign-off. Several jurisdictions now require disclosure of AI use in court filings specifically, following a run of cases involving fabricated citations. Professional liability generally attaches to the human regardless, which is a strong practical argument for checking everything.

Commissioned creative work. If a contract specifies original work, generated content may breach it. Many creative contracts now address this explicitly. Check before delivering, not after.

Photography presented as documentary. News, evidence, insurance claims, scientific publication. Manipulation that changes what the image asserts about reality is a different category from adjusting exposure, and generated imagery in these contexts is not a disclosure issue but a prohibition. This is why our terms rule it out for the Photo Humanizer.

Platforms with policies. Some publishers, marketplaces, and stock libraries require labelling. Some jurisdictions are moving toward mandatory labelling of synthetic media. These change quickly; check the specific platform.

Where it is genuinely optional

Routine commercial content. Product descriptions, help documentation, meta descriptions, internal summaries. Nobody's decision changes. Disclosure here is noise.

Your own correspondence. Email, messages, notes. It is your communication and the assistance is your business.

Work where the output is what is being judged. If someone is buying a functioning marketing page and the page works, the drafting method is not material — unless they specified otherwise.

The consistent principle: where the audience is buying the result, method is usually immaterial. Where they are buying provenance — that a specific person did this thinking, that this photograph records a real moment — method is the whole thing.

How to write a disclosure that works

Bad disclosure is worse than none, because it either says nothing or reads as an apology.

Be specific about what the tool did. "This article was created with AI" tells the reader nothing about what to trust. "The first draft was outlined with Claude; the benchmarks, analysis, and conclusions are mine, and every figure was verified against the source data" tells them exactly where to place their confidence.

Put it where it will be read. A note at the top for anything substantive. Buried footers are technically disclosure and practically concealment.

Do not apologise. Using a tool competently is not a confession. Defensive framing invites the reader to be suspicious of a workflow that did not warrant it.

Say what you verified. This is the part readers actually want. Fabricated facts are the real risk, so stating that you checked is more reassuring than any statement about the tool.

Some workable examples:

Drafted with AI assistance. All data, quotes, and technical claims were independently verified by the author.

I used a language model to help structure this piece and tighten several sections. The argument, the research, and the errors are mine.

This image was generated, not photographed.

Portions of this documentation were drafted with AI assistance and reviewed by the engineering team.

Each is short, specific, and unembarrassed.

Building a policy for a team

If you publish at any scale, individual judgement will produce inconsistency. A short written policy prevents most of it. Three things to settle:

Which tiers are permitted for which content types. Language assistance everywhere; substantive generation permitted for product documentation but not for research reports; generated imagery never for anything documentary.

Who is accountable. Name the human responsible for accuracy in every case. The most important effect of AI use policies is not restricting tools — it is making sure a person owns the output.

What gets labelled and how. Pick standard wording and use it consistently. Inconsistent labelling reads as arbitrary and undermines the labels that matter.

Publishing that policy is itself worth something. It answers the question before anyone asks it, and it signals that the decisions were considered rather than improvised. Ours is on our editorial standards page, including how we use these tools in our own writing.

The honest bit about tools like ours

We build a Text Humanizer that rewrites prose to read more naturally. It would be convenient for us to argue that disclosure never matters. It does matter, in the contexts above, and using a rewriting tool to conceal work you did not do is against our terms and a bad idea besides.

The distinction we draw is between assistance with expression and substitution of authorship. Reworking how your own argument reads is editing, and editing has never required disclosure. Passing off an argument you did not make is a different act, and no amount of rewriting makes it not one.

What this comes down to

Disclosure is not about the technology. It is about not letting someone believe something false about how work was produced when that belief matters to them.

Most of the time it does not matter and disclosing is clutter. Sometimes it matters enormously and omitting it is deception. The audience-decision question separates those two cases reliably, and once you have separated them, the actual disclosure is one sentence.