Article50.io
Method · Checks · Limits

How the scan works

Last updated:

How the scan works, in short: Article50.io loads one public page in a real Chromium browser with scripts running, recognises AI chat, voice and avatar widgets, looks for an AI disclosure in seven languages, measures how clearly it is shown, and reads AI provenance markers in images. The free scan shows the most severe finding, never a legal verdict.

Key point: the full report says what was checked and what could not be checked. Below, step by step, is how the scan gets there, and what it can't check.

Step 1: Your page loads in a real browser

We open the URL you give us in a real Chromium browser, not a simple HTML download. Scripts run and widgets load, at a desktop screen size (1366 × 768). Then we wait up to 5 seconds for chat widgets that load late, and stop sooner once a known chat widget has appeared on screen. To keep scans quick, the browser doesn't download images, fonts or media; images are read separately in Step 6. No scan runs in the browser for longer than 45 seconds.

We scan the page at the address you give us; we don't crawl the rest of the site.

Step 2: We look for AI chat, voice and avatar widgets

We recognise widgets from known vendors by their scripts, the requests the page makes to their loader servers, page globals, element IDs and, for voice and avatar widgets, the embed markup. Vendors we currently recognise:

Chat: Intercom, Drift, Zendesk, Crisp, Tidio, LiveChat, HubSpot chat, Freshchat, tawk.to.
Voice: ElevenLabs voice agent, Vapi voice agent, Telnyx AI agent.
Avatar: D-ID avatar, Tavus avatar, Synthesia presenter video, HeyGen video.

A custom-built chatbot that doesn't match a known vendor may not be recognised. It is only picked up if its markup uses common names such as "chatbot", "ai-chat" or "chat-widget", and is then reported as a custom chat widget. Voice and avatar widgets are recognised from the vendors above only.

Step 3: We read the text visitors see, including inside widgets

We read the visible text of the page. For chat, voice and avatar widgets we also read text inside chat-like frames (among the first 25 frames on the page) and chat-like shadow roots, which ordinary page text doesn't include. Frames that don't look like chat widgets aren't read, and the reading is best effort and time-boxed.

We read only what a widget shows as it first loads. Nothing is clicked or opened, so a disclosure that appears only once a conversation starts isn't seen. How a disclosure inside a frame from another site is presented can't be measured, and the report says so.

Step 4: We look for an AI disclosure, in 7 languages

We recognise wording that tells visitors they are talking to AI in English, German, French, Spanish, Italian, Dutch and Portuguese. On the page, only sentences addressed to the visitor count, such as "You're chatting with an AI assistant". Marketing lines like "Powered by AI", bare labels like "virtual assistant", and negated phrases ("you are not talking to an AI") don't count as a disclosure there.

Inside a chat widget, where the text is the bot speaking, more wording counts: first-person statements such as "I'm an AI assistant" and labels that name the bot as an AI. Negated phrases still don't count, and in the six languages other than English a label has to say AI.

When we find a disclosure, we compare its language with the page's language setting where both are clear, and say "couldn't determine" otherwise. Labels on AI-generated content and on images are recognised in English only. Avatar labels are recognised in all seven languages. Read more in AI disclosure and language.

Our non-English phrase lists are still being reviewed by native speakers. If we miss a correct disclosure in your language, tell us at [email protected] and we'll fix it.

Step 5: We check how clearly a disclosure is presented

Where we find a disclosure, we measure how it is presented as the page first loads. We report measurements and flag only clear problems:

  • →Visibility: whether the text is hidden (display or visibility off, zero size, off-screen, near-zero opacity, clipped, fully transparent text) or very small (under 10px).
  • →Readability: the contrast of the text against its background and its font size. Contrast is shown against 4.5:1, a common technical benchmark and not a legal threshold, and only a ratio below 3:1 is flagged. Text under 10px and contrast below 3:1 are scanner heuristics, not an Article 50 pass or fail.
  • →Assistive technology: whether the text is hidden from assistive technology (aria-hidden, inert, the hidden attribute, or not rendered). This reads attributes and rendering state. We don't run a screen reader or an audio test. The Commission's guidelines (paragraph 144) say Article 50 imposes no distinct or additional accessibility requirements. Other accessibility law can still apply.
  • →Language: whether the language of the matched text differs from the language set for it on the page. It is flagged only when both are clear. A page marked English with text clearly in another language is "couldn't determine", because English is often just a default.

Anything we couldn't inspect, such as gradient or image backgrounds, text in a frame from another site, or text we couldn't locate, is reported as "couldn't determine" and never produces a finding.

Step 6: We check images for AI provenance markers

After the browser is closed, we read the start of up to 8 images on the page (the first 256 KB of each, within a shared 5-second limit). We pick the social-share image first, then images by size, and skip icons, logos, SVG, video and tiny images. We look for provenance data that says an image was made by a trained model: an IPTC digital source type in the image's XMP metadata, or in a C2PA (Content Credentials) manifest.

If an image says it is AI-generated and no label, caption or alt text near it, or text elsewhere on the page, says so (English wording only), we report it. That is what the check looks for. A provenance marker does not by itself show the image is a deep fake, and alt text or unrelated page text is not necessarily a perceptible disclosure. The flag is conditional and needs a person's review of legal scope. It does not decide that a disclosure is legally sufficient. A missing marker is never a finding, because many AI images carry none, and markers are easy to strip.

Separately, if the page's own markup says content is AI-generated (a generator tag naming a known AI tool, or an explicit AI-generated attribute) and no visible text discloses it, we report that too.

Step 7: We never show a blocked page as clean

If the page redirects to a different site, shows a bot check or security challenge, answers with an HTTP error (400 or above), loads with almost no visible text (under 50 characters), or fails to load or times out, we say we couldn't scan it. We never show that as a clean result. A redirect within the same site, including to or from a "www" or subdomain address, isn't treated as another site. For a redirect to a different site, we show where it went and let you scan that address.

What you get

See a sample report, or run a free scan. For the free scan and €499 report pricing, see the pricing page.

Check these by hand

The scan does not run these tests. Open the widget, the call or the image yourself:

  • →Open the chat and read the first turn. A missing notice in the scan's first snapshot is a possible gap, not proof the duty was missed: the notice may lawfully appear at the first interaction after the widget opens.
  • →For most chats, one prominent notice at the start of the session is likely to be enough. Longer or immersive chats, sensitive advice, vulnerable users and AI companions may need periodic or context-aware reminders. There is no fixed interval. The Commission's guidelines (paragraphs 37 and 40, C(2026) 7702 final, 5 November 2026) describe this. They are non-binding and do not add a new November obligation. Guideline paragraph numbers on this page refer to Commission Guidelines C/2026/6005 (OJ C, 12.11.2026), adopted as C(2026) 7702 final on 5 Nov 2026, non-binding (point 5).
  • →Check handoffs between a bot and a person, sessions that resume, and what the system answers if you ask whether it is AI.
  • →On a voice call, keep the opening spoken disclosure, and repeat it as appropriate on a longer call, after an interruption, when the interaction resumes, or when the agent's role changes. The scan cannot hear audio.
  • →If an image carries an AI provenance marker, decide yourself whether it is a deep fake. Appreciable resemblance and a false appearance of authenticity in context are both required. Photorealism alone is not enough. Alt text or unrelated page text is not necessarily enough for a perceptible disclosure.

What the scan can't check

A clean result is not a compliance clearance, and this page is not legal advice. For the rules themselves, see the Article 50 guide.

For how a live-page scan compares with the Commission's free AI Act questionnaire and other tools, see EU AI Act checkers: what an Article 50 check should cover.

Frequently Asked Questions

Does the scan use a real browser?

Yes. Your page is loaded in a real Chromium browser so scripts and widgets run as they would for a visitor.

Which languages does it recognise?

AI disclosures addressed to visitors in English, German, French, Spanish, Italian, Dutch and Portuguese. Content and image labels in English only.

Does a clean result mean my site meets Article 50?

No. It means the scan found no obvious issue in what it could check. The full report lists what it could not check.

What happens if my site blocks the scan?

We tell you we couldn't scan it. We never show a blocked or empty page as clean.

Does the scan check images for AI?

It reads provenance markers some AI tools embed in images. It does not guess from how an image looks, and a missing marker is never reported as a problem.

Check your site automatically

Article50.io is an automated scanner for the Article 50 transparency rules of the EU AI Act. It scans a website for potential Article 50 transparency gaps and generates a report of what the scan found and what it could not check, with remediation guidance, implementation instructions and disclosure wording you can paste in.

The free scan shows your single most severe finding. Scans are capped at 45 seconds — no signup, public pages only.

Related Article 50 resources

Automated technical guidance, not legal advice. Citations refer to Regulation (EU) 2024/1689.