When You Actually Have To Disclose That You Used AI
The question arrives eventually: do I have to say I used AI for this? Usually it arrives late, when the document is written and someone senior asks.
The answer is not the same for a rewritten email and a client deliverable, and treating them alike leads either to pointless disclaimers on everything or to an uncomfortable conversation about something you should have flagged.
Start With What You Actually Agreed To
Before reasoning from principle, check whether the answer already exists.
Client contracts increasingly contain AI clauses, and they are specific. Some prohibit use on the engagement outright. Some require notice. Some restrict which categories of tool are acceptable, typically by whether data leaves your control. AI clauses in client contracts and what they mean covers the common shapes.
Employers have policies, and grant applications, journals, courts, and public tenders increasingly have their own rules. Transparency obligations in the EU AI Act apply to particular categories of system rather than to your email, but they are shifting what people expect by default. Where a written rule exists, it governs, and no amount of reasoning about whether it should replaces reading it.
Most disclosure failures are not judgement calls. They are people who did not know a clause existed.
The Useful Test Is About Substance, Not Assistance
Where no rule applies, the question that produces sensible answers is not “did AI touch this” but “would the reader’s assessment change if they knew”.
Rewriting your own paragraph to be clearer does not change what the reader is assessing. The judgement, the facts, and the commitments are yours; the tool adjusted the phrasing. Nobody expects a footnote for spellcheck, and this is closer to that than people assume.
Generating substantive content you then lightly reviewed is different. The reader is assessing analysis, and the analysis did not come from you. That is the case where non-disclosure starts to mislead.
The line sits at whose judgement produced the substance. It is not a bright line, but it is a workable one, and it explains why the same tool can be unremarkable in one document and disclosable in another.
Three Situations Where You Should Just Say So
Some cases are clear enough to treat as rules.
When you are being evaluated on the writing itself. A writing sample, an application, an assessed exercise. The thing being measured is your ability to produce the text, so using a tool changes what the assessment measures.
When you cannot personally vouch for every claim. If AI produced factual content you have not independently verified, disclosure is the honest position, and the better move is usually to verify rather than to disclose.
When someone has asked. If a client or employer asks directly, the answer is the truth, and a hedged answer is worse than either a yes or a no. How to answer “did you use AI” in a proposal covers the phrasing for the case that comes up most.
Say It Plainly Or Do Not Say It
Bad disclosure is worse than none. It reads as either a boast or an apology, and both undermine the work.
Before:
Please note that this document was created with the assistance of AI tools. While every effort has been made to ensure accuracy, AI-generated content may contain errors and should be independently verified.
After:
I drafted this and used a rewriting tool to tighten it. The analysis and figures are mine and I have checked them.
The first disclaims responsibility for your own document, which is precisely the opposite of what a disclosure should do. It invites the reader to distrust everything and tells them nothing about what was actually done.
The second states what happened, keeps ownership of the substance, and takes about four seconds to read. If you cannot write the second version honestly, that is useful information about whether you should be sending the document.
A Wrivio Context for disclosure statements could say:
Rewrite this as one or two plain sentences stating what was done. Keep responsibility with the author. Do not add disclaimers, hedging, or language suggesting the content may be unreliable. Keep every fact exactly as written.
Press Ctrl+Shift+Space, paste your draft, and check the diff. Models are strongly inclined to add liability language here, and that is the specific thing to reject.
The Tool You Choose Changes The Conversation
One practical point that gets missed. Much of the anxiety about AI at work is not really about authorship, it is about confidentiality: whether client material was pasted into a service that retains it.
If your tool processes text on your own machine, that concern does not arise, and the conversation becomes a narrower one about authorship alone. That is a much easier conversation, and it is worth being able to say plainly which category you are in.
Which tasks should stay local is the sorting exercise, and it is worth doing before someone asks rather than after.
This is a description of common practice, not legal advice. Where a contract, regulator, or professional body governs disclosure, get advice on the specific obligation.
Common Questions
Do I need to disclose using AI to rewrite my own email?
Almost never. The judgement, facts, and commitments are yours and the tool adjusted phrasing, which is closer to spellcheck than to authorship. Check for a contract clause or employer policy that says otherwise.
When is disclosure genuinely required?
When a contract, policy, or institution requires it; when you are being assessed on the writing itself; when AI produced substance you cannot personally vouch for; and whenever someone asks you directly.
What is the right way to word a disclosure?
One or two plain sentences saying what was done, keeping responsibility for the content with you. Long disclaimers that warn the reader the content may be unreliable undermine your own document.
Does it matter which AI tool I used?
For authorship, not much. For confidentiality it matters a great deal, because a tool that processes text on your own machine removes the question of whether client material reached a third party.
Download Wrivio for Windows to keep confidential drafting on your own machine, so the only question left is about authorship.
Read Next
What AI Detectors Can And Cannot Prove
Detection scores are being used to make decisions about people. What the tools actually measure, why they misfire on specific groups, and how to respond to an accusation.
Content Credentials Are Coming for Text: What Provenance Marking Means for Writing
The C2PA standard now covers text, and Article 50 requires machine-readable marking of synthetic content. What that actually means for the email you just rewrote.
The EU AI Office Can Now Enforce: What Its Powers Actually Are
From 2 August 2026 the Commission can investigate general-purpose AI providers, demand model access, and fine them. What that means for the companies you buy AI from.
Why Two Leaderboards Rank The Same Model Differently
One board has a model first, another has it fifth, and both are honest. The four reasons scores diverge, and which board to trust for writing work.
This article is filed underPrivacy & Compliance, which has 53 articles.