I couldn't say in one sentence what to tell users when we used AI. So I built a drafter.

We shipped a feature that used a model to summarize support tickets. Quietly. And then a reviewer asked the only question that mattered: when a user reads the output, are we telling them it was generated by AI? I didn't have an answer, and the EU AI Act expects one.

Articles 49 and 50 of the Act are about exactly this — telling people when AI is involved, whether that's a general-purpose model's capabilities or AI-generated content a user might mistake for human-made. The obligation is concrete wording, aimed at the user, at the point of interaction. I kept getting it wrong because "disclose that AI was used" is a sentence, not a disclosure.

The duty isn't "mention AI somewhere." It's specific: what was used, that the output was AI-generated or manipulated, and enough that a person can make an informed decision. DiscloseKit takes the facts of your situation — what model, what surface, what the user sees — and drafts the disclosure text those articles call for.

It's decision-support, not an oracle. The hard part, deciding which article even applies to your case, is something you reason through with the tool; the tool then turns that reasoning into wording you can actually paste near the interaction.

I'm not anti-model. Drafting readable disclosure prose from a set of facts is genuinely open-ended — that's where a model earns its place, turning "we use a summarization model on tickets" into something a user understands in two seconds. The part that must not be open-ended is the article mapping: which duty, which surface, what must be stated. That stays anchored to the written text of the Act.

DiscloseKit is $19/mo. If your product uses AI anywhere a user might not expect it, draft the line you'll show them before a regulator asks why you didn't.