For teams shipping AI content or features. EU AI Act Art. 50 transparency duties have applied since 2 Aug 2026 (fines up to EUR 15M or 3% of global turnover), and deepfakes must be labelled even without intent to deceive. Not legal advice — a gap finder.
Free with no sign-up, and the check runs inside your browser, so what you enter never leaves your device. The result is a signal to help you decide, not proof for money or legal decisions.
If you ship AI features or AI-made content, disclosure duties probably reach further than you think. Use the checklist above to find gaps, then read on for what counts as disclosable and where the marking has to sit. This is a gap finder, not legal advice.
Four buckets cover most of it. First, AI that talks to people: a chatbot or AI agent has to make clear the user is not speaking to a human, at the first interaction, unless it is already obvious.
Second, generated or manipulated image, video and audio. Third, deepfakes of real people — and this is where teams get caught out: disclosure applies even with no intent to deceive, including material made as a joke. Fourth, AI-written text on matters of public interest, typically news-style content.
Artistic and satirical work can qualify for a lighter form of disclosure that does not spoil the piece. Lighter means the placement can change, not that the disclosure can disappear.
Disclosure has two layers: one a person reads, one a machine reads. On-screen wording is the first; a signed provenance record, watermark or metadata embedded in the file is the second. Having only one of the two is half the job.
The human-readable layer cannot hide. Not behind a "more" link, not in the page footer. It belongs where the content is, at body text size, in wording nobody has to decode. Tucked between hashtags does not qualify.
The machine-readable layer is usually written by the generation tool itself. The catch is how easily it is lost: editing, re-compression and screenshots all strip it. Check the final file right before you publish. And when uploading to a platform, switch on its AI-label option too. That runs separately from anything inside the file.
The EU AI Act transparency duties apply to providers and deployers serving the EU market. Where your company sits matters less than who your users are, and being small is not an exemption. The duties have applied since August 2026, and the penalty ceiling is not trivial.
Korea brought its own AI framework legislation into force in January 2026, carrying a duty to make clear when output is generated by AI. The wording differs from the EU text, but the direction is the same — legible to people, and readable by machines.
In practice, naming an owner changes more than reading the law does. Missing labels usually come from nobody checking rather than nobody knowing. Put the disclosure items on your release checklist and give them an owner. The checklist on this page works as a first draft.
The first reply on the contact screen was automated, yet it carried an agent name and a profile photo. Handover to a human came several turns later. Dressing the bot up as a person raises the need for disclosure rather than lowering it. A single line above the first reply, plus a visible route to a human, settles it.
Real employees, filmed normally, were composited against a city skyline that does not exist. The people and their words were genuine. Genuine people do not exempt a scene that has been made to look unlike reality. Put the on-screen notice and the in-file marking in place, and switch on the platform AI label at upload. That is what fills both layers.
A team decided that output from an internal meeting-summary tool was disclosable too, and began stamping a notice on every internal document. This checklist is built to find gaps, so it will not push back when you overreach. Marking output that never leaves the company makes the notice ordinary, and an ordinary notice stops carrying signal where it matters. Define the external boundary first and label to that line.
Ordinary brightness and colour correction is not the target. The test is whether a real scene has been made to look materially unlike itself. Adding people or objects that were never there, or changing an expression or what someone appears to say, sits on the disclosure side regardless of how small the edit was.
The transparency duties aim at content provided or published to users. Output from an internal tool that never leaves the building is not the focus. The moment it is folded into external material it becomes disclosable, so write the boundary down rather than leaving it to judgement.
No. It finds gaps; it does not certify anything. Real assessment depends on how your service is built and which markets it serves, so get professional review if the stakes are meaningful.
Nobody, unless the contract says so. Commissioning the work does not move the duty: the party publishing the output is still the one publishing it. Put disclosure of generative tool use and preservation of machine-readable marking into the delivery terms, and add a step that checks the delivered files actually still carry it.