Disclosing that you used it

E6·L7 covered labelling generated media. This is the same question for the work you do every day — the report, the analysis, the code, the application.

The answer is neither “always” nor “never”, and the middle is more workable than either.

The test

Would the reader change how they weigh this if they knew?

If yes, tell them. If no, you are describing your writing process, which nobody needs.

Nobody discloses spellcheck, because knowing changes nothing about how a reader should treat the text. At the other end, an assessor grading your understanding weighs the work entirely differently if it was produced rather than reasoned.

Most cases sit between, and the test sorts them.

When it is always yes

Assessed or credentialed work, where a rule exists — and where none does, ask rather than assume.

Where a rule already applies. Publishers, journals, employers, clients, courts, competitions. Increasingly the law: the EU AI Act’s transparency obligations are the most cited, and more are arriving.

When someone is relying on your expertise specifically. They engaged your judgement. If a substantial part came from elsewhere, that is material to what they are paying for.

When it affects how they should check it. “The figures came from a model” is information a reader needs to know how much verification is theirs to do.

When it is usually no

Routine correspondence. Formatting and tidying. Brainstorming that you then did something with. Code you fully understand and have tested. Research you verified independently.

The common thread: the assistant helped you produce something you can defend, and knowing would not change how anyone should treat it.

Say what it did

The most common mistake is a vague, slightly apologetic admission — “AI was used in preparing this document” — which is worse than nothing. It invites suspicion about everything without telling anyone what to actually distrust.

Be specific:

First draft outlined with an assistant; all figures verified against the source data; conclusions and recommendations are mine.

Code generated with assistance; tests written by hand and all cases reviewed.

These read as competence, because they are. They tell the reader exactly where to place weight and what was checked — which is E7·L8’s uncertainty formula applied to process rather than to facts.

Norms are unsettled, so decide yours

There is no stable convention yet. It differs by field, by organisation, and by year, and anyone claiming otherwise is describing their own corner.

Given that, the asymmetry argument from E6·L7 holds here too. Disclosing costs a line. Being found not to have disclosed costs the credibility of everything else you have produced — including the parts that were entirely yours. That is not close, and it does not require the norms to settle first.

Try it now (5 minutes)

Take something you produced with assistance. Apply the test.

If it needs a line, write it — specific, naming what was assisted and what you verified. Notice it reads better than silence would have.

Check your understanding

1. The disclosure test is: Whether an assistant was used at all Whether the reader would change how they weigh it if they knew Whether more than half the text was generated

2. Why is "AI was used in preparing this document" a poor disclosure? It is too long It invites suspicion about everything without saying what to distrust It admits more than necessary

3. The argument for disclosing even when nobody requires it: It is always legally safer The cost is a line; being found out later damages everything else you produced Readers prefer AI-assisted work

Recap

Disclose when knowing would change how a reader weighs the work, always for assessed work and where a rule or reliance on your expertise applies. Say what the assistant did and what you verified rather than making a vague admission — and remember the asymmetry: a line of text against the credibility of everything else you have written.