Free scanner + open vendor dataset for the AI-disclosure rule that has applied since 2 August 2026.
👉 Run a scan — no signup, nothing stored, runs in your browser.
Most EU AI Act tooling is built around high-risk (Annex III) obligations. The Digital Omnibus, given final Council approval on 29 June 2026, deferred most of those to 2 December 2027.
Article 50 transparency was not deferred. It has been binding since 2 August 2026:
| Date | What applies |
|---|---|
| 2 Aug 2026 | Art. 50(1) — people must be informed they are interacting with an AI system, unless obvious. Art. 50(4) — deepfake and public-interest AI text disclosure. |
| 2 Dec 2026 | Art. 50(2) — machine-readable marking of synthetic audio/image/video/text. Four-month transition for systems already on the market before 2 Aug 2026. |
| 2 Dec 2027 | Most standalone high-risk obligations (deferred by the Digital Omnibus). |
If you enabled an AI agent in Zendesk, Intercom, Gorgias, Tidio, Crisp or HubSpot, you are the deployer and the duty is yours, not your vendor's.
- Fetches the public HTML of a URL.
- Fingerprints known conversational-AI vendors (36 and counting).
- Searches visible page copy for AI-disclosure wording in 10 EU languages.
- Reports per-article findings with an explicit confidence level.
It deliberately does not claim you are non-compliant. It reports what the page source does and does not show, and what that means you need to confirm.
- Homepage-only unless you pass a deeper URL. Widgets usually live on
/contactor/help. - Cannot see widgets injected later by a tag manager.
- Cannot open your chat widget or read your vendor console, so it cannot tell whether your AI mode is switched on.
- Heuristics, not an audit. Not legal advice.
data/vendors.json — CC-BY-4.0. Each vendor carries:
aiNature is what drives the Article 50(1) reasoning:
- ai-native — the product is an LLM agent by design. The duty applies.
- ai-optional — vendor ships both scripted and AI modes. The duty applies if AI mode is on; only the operator can confirm.
- rule-based — scripted flows. Art. 50(1) can still bite where the interface reads as human.
PRs adding vendors are welcome. Add the fingerprint, the aiNature, and where in that vendor's console the disclosure actually goes.
node scanner/scan.mjs example.com
node scanner/scan.mjs @sites.txt > results.jsonlOutputs JSONL. The CLI and the browser app import the samescanner/core.mjs, so verdicts are identical by construction.
Full write-up and open data: https://clearlabel.eu/study/
Scanned 883 EU sites on 16-17 Aug 2026, 667 of them readable, across two sampling frames:
| Frame | Scanned | Readable | Runs a chat/AI widget | Of those, no disclosure |
|---|---|---|---|---|
| Tranco top EU ccTLDs | 703 | 532 | 4.9% | 92.3% |
| Independent EU online shops | 180 | 135 | 24.4% | 97.0% |
| Combined | 883 | 667 | 8.8% | 94.9% |
Widget density is five times higher among online shops than across general EU sites, so the Tranco figure is a floor, not a market estimate. Only 0.4% of all readable sites carried any machine-readable AI-content marking ahead of the 2 December 2026 deadline.
Raw per-domain results in study/.
ClearLabel provides structured information and document drafts. It is not a law firm and this is not legal advice. Penalties under Art. 99(4) are set nationally and capped at €15m or 3% of worldwide turnover. For a binding view on your exposure, consult a qualified adviser in your member state.
{ "id": "crisp", "name": "Crisp", "aiNature": "ai-optional", // ai-native | ai-optional | rule-based | content-gen"aiProduct": "MagicReply / Crisp AI", "patterns": ["client\\.crisp\\.chat", "\\$crisp"], "disclosureHook": "Operator nickname + welcome message in Crisp settings" }