Disclosure: when to say a machine was involved

"Should we say AI was involved?" gets treated as one question. It bundles several: contexts where a rule already exists, contexts where a relationship implies one, and contexts where disclosure serves nobody.

Where the rules already exist

Do not invent a position in these contexts. One has been settled.

Academic and scientific publishing. The major bodies converge: an AI cannot be an author, because authorship carries accountability and a model cannot be accountable. Writing assistance is disclosed in acknowledgements; analytical or data-generating use belongs in methods. Non-disclosure can be treated as misconduct. ICMJE, COPE and the large publishers agree on this, so there is nothing for an individual organisation to decide.

Education. Institutional policies vary and UNESCO's guidance is the common reference point. The rule that matters: the student's own institution governs, and "my employer allowed it" is irrelevant there.

Regulated professional advice. Sector rules on how advice is produced and who is accountable for it apply unchanged. Check yours; do not reason from first principles.

Products and interfaces. Transparency obligations under the EU AI Act attach to certain AI systems interacting with people. That is a product question rather than a document-authorship one, and it needs looking up for your specific case.

Everywhere else: a workable test

Outside settled contexts, disclosure is a judgement. A pair of questions resolves most of it.

Would the recipient's decision change if they knew? A client choosing a supplier partly on craft, a reviewer assessing whether you understand the subject, a reader weighing a personal recommendation — for these, involvement is material and silence is a small deception.

A build script, a formatted spreadsheet, a meeting summary you checked: the recipient wants the artefact to be right, and how it was produced does not change what they do next.

Are you claiming something the process undercuts? "I read all forty submissions" after skimming a generated summary is a false claim about your own effort, regardless of policy. Disclosure is not the fix here; the fix is not making the claim.

What a disclosure should say

The useless version names a tool. The useful version names the division of labour.

• Weak: "This report was created with AI assistance."
• Better: "Drafted with AI from our source data; figures verified against the originals; conclusions and recommendations are ours."

The second tells a reader what to trust and what was checked, which is the information they actually wanted. It also reads as confidence rather than apology, which is closer to the truth.

Blanket disclosure ages badly

A footer on every document saying AI may have been involved conveys nothing after the third document, and it will read oddly within a year or two — the way a note about word processors would now.

Disclose where it is material or required. Elsewhere, the standing expectation that people check their own work does the work, and clause 4 of a decent acceptable-use policy already says so.

What goes wrong

Treating one answer as covering every context. Journal submission and an internal summary are governed by different things.

Disclosing to transfer responsibility. Saying a machine helped does not move accountability for accuracy. Clause 4 exists precisely because it does not.

Disclosing everything. The signal disappears and the useful cases are buried with the trivial ones.

Naming a tool instead of a division of labour. "Made with X" tells the reader nothing they can act on.

Assuming your policy overrides someone else's. A student's institution, a journal, a client contract and a regulator each govern their own context.

How to check it worked

Take three recent pieces of work and apply both questions. If all three come out the same way, the test is probably being applied mechanically rather than thought about — the interesting cases are the ones where reasonable people would split.