Instant AI Act transparency check

Use the box above. You'll get two clearly separated answers: our independent verdict if we've hands-on reviewed that agent, and a live surface read of its Article 50 signals either way. We never merge those two into one number — tested and observed are different things.

Fair — and you don't have to. The check is anonymous about you: we see a URL, never who asked, so "Company X uses Tool Y" is something we structurally cannot learn.

If you'd rather not type it at all, work from the category instead. Your obligations as a deployer depend mostly on what the tool does, not which brand it is:

• A chatbot / assistant your customers or staff talk to → people must be able to tell they're dealing with an AI, unless it's obvious from context. Applies 2 Aug 2026, no grace.
• A generator of images, video, audio or synthetic text you publish → deepfake and synthetic-content labelling duties apply. Applies 2 Aug 2026, no grace.
• Anything producing AI content you republish → check whether the provider marks output in a machine-readable way; the marking duty has a grace period to 2 Dec 2026 for systems already on the market.

Then build the paperwork with the Evidence File — it works from categories too.

We won't rate it — and that's deliberate. Hlido's ratings mean something because every one of them is independently tested from the outside. A private internal agent can't be tested that way, so publishing a Hlido score for it would quietly break the thing that makes our scores worth reading.

What we can do is hand you the same test dimensions to run yourself, and keep the result honestly labelled as self-reported:

1. Interaction disclosure — when a person interacts with your agent, are they told it's an AI? Where exactly does that appear (first message, UI label, onboarding)?
2. Synthetic output labelling — if it produces images/audio/video or published text, is that output labelled as AI-generated?
3. Machine-readable marking — is generated output marked in a machine-readable way (signed/timestamped metadata, watermark) so downstream systems can tell?
4. Detection & verification — can anyone verify that marking, and is there a documented way to do so?
5. Evidence — for each of the above: a screenshot, a doc link, or a config reference, dated.

Record your answers with evidence and a date, keep it on file, and re-run it whenever the agent changes — the same discipline we apply to the public corpus. The Evidence File gives you a dated document to attach it to.

Want an independent party to test a private agent under NDA rather than self-assess? That's a lab engagement, not a public rating — tell us what you're running.