Question · Art. 50(1), 50(2), 50(4)

Do AI Avatars Require Disclosure Under Article 50?

Last updated: 9 November 2026

AI avatars require disclosure in most cases, often under several rules: an avatar that talks with people needs the Article 50(1) interaction notice, a realistic avatar that could pass as a real person can be a deep fake needing disclosure under Article 50(4), and the avatar vendor marks output under Article 50(2). Article50.io scans one public page for free.

Which obligations apply depends on what your avatar does and how real it looks. Below: three questions to ask about your avatar, when a realistic avatar counts as a deep fake, and common avatar setups.

Scan for free

Three questions to ask about your avatar

Does it talk with people? → 50(1)

If people interact with it directly, they must be informed it's an AI system unless that's obvious. A lifelike face and voice make "obvious" much harder to argue. More on 50(1) →

Is it generated media? → 50(2)

The provider of the system generating the avatar's video, image or audio must mark that output in a machine-readable format. That's your avatar vendor — or you, if you built the generator. More on 50(2) →

Could it pass as real? → 50(4)

A realistic avatar can be a deep fake only if the criteria are all met: appreciable resemblance and a false appearance of authenticity in context. Photorealism alone does not decide it. If they are met, the deployer — whoever uses it in a professional capacity — must disclose that it is artificially generated or manipulated. More on 50(4) →

Read the full text of Article 50 on the Commission's AI Act Service Desk.

When a realistic avatar counts as a deep fake

Article 3(60) defines a deep fake as AI-generated or manipulated image, audio or video content that resembles existing persons, objects, places, entities or events and would falsely appear to a person to be authentic or truthful.

A digital replica of a real person is an example of a "person" in paragraph 113. It does not, by itself, qualify as a deep fake. Invented people are not an easier case. The Commission's guidelines, adopted on 5 November 2026 as C(2026) 7702 final, read "existing" to include subjects that exist, can plausibly exist or could plausibly have existed, and treat realistic AI-generated human avatars and personas as persons for that criterion (paragraph 113). The 20 July 2026 text (C(2026) 5054 final) was a draft submitted for approval and said the same. Paragraph 113 says the four criteria are cumulative. Paragraph 114 says a false appearance of authenticity or truthfulness is assessed in context, and that photorealism alone is not determinative. A photorealistic avatar of an invented person therefore does not automatically fall within the definition.

Two caveats. The guidelines are non-binding (paragraph 5) — only the Court of Justice can give an authoritative interpretation of the Act. And the text of Article 3(60) itself says "existing", so the guidelines' reading of plausible existence is not a court ruling. Review resemblance and whether, in context, the avatar would falsely appear authentic. A disclosure can still be a cautious practical step. That step is not a finding that the avatar is in scope.

A clearly stylised, cartoon or illustrated avatar is unlikely to "falsely appear to be authentic", so it generally won't be a deep fake — though if it chats with people, the 50(1) interaction disclosure still applies.

Where the avatar is part of an evidently artistic, creative, satirical or fictional work, the 50(4) disclosure can be limited to flagging that generated content is present, in a way that doesn't hamper the work.

Common avatar setups

Setup	Likely rules	What it means
Photorealistic avatar that holds live conversations (support agent, video-call assistant)	50(1) · 50(4) · 50(2)	Tell users they're talking to an AI, and disclose the avatar is artificially generated. The avatar vendor marks the output.
Cartoon or clearly stylised mascot that chats with visitors	50(1)	Interaction disclosure still applies unless it's genuinely obvious it's AI. Unlikely to be a deep fake, as it wouldn't pass as authentic.
Pre-rendered AI presenter in a marketing or training video	50(4) · 50(2)	No live interaction, so 50(1) does not apply. Article 50(4) disclosure may apply when the presenter meets the deep-fake definition; provider marking duties are separate. Article 50 for video creators.
Digital replica of a real person (your CEO, a spokesperson, a celebrity)	50(4) · 50(2)	Squarely within the deep fake definition. Add 50(1) if the replica also converses with people.

Whichever rules apply, visible disclosures must be clear and distinguishable and given at the latest at the first interaction or exposure — a watermark in the file isn't something a viewer can see. If your avatar is really a chat widget with a face, start with does my chatbot need an AI disclosure?

Which scanner checks AI avatars for Article 50?

Article50.io's scan checks web pages for EU AI Act Article 50: it loads your page in a real browser and recognises the D-ID avatar, Tavus avatar and Synthesia presenter video widgets by their scripts and embed markup. It reads the text visitors see, plus the text the widget shows as it first loads, inside chat-like frames among the first 25 frames on the page (including D-ID frames served from that vendor's own agent host) and open chat-like shadow roots, within a short time limit. Tavus is a custom element, so its first-load text is read from an open shadow root on that element when the browser exposes one. A Tavus frame from a per-conversation address is not on the frame-address list, so text that exists only inside that frame may not be read. Then it checks for a label telling visitors the avatar is AI. Avatar and presenter labels are recognised in all seven supported languages: English, German, French, Spanish, Italian, Dutch and Portuguese. Labels on images are recognised in English only.

What it can't see:

Avatars from other vendors: only those three vendors' avatar widgets are recognised.
Text inside a Synthesia presenter frame (the embedded share player). The widget is detected, but text inside that frame isn't read, so a label shown only there won't be found. Put the label on the page next to the video.
How a disclosure inside a D-ID or Tavus frame from another site is presented (size, contrast, position): the scan reads the text but can't measure how it is shown. (Synthesia frames: see the bullet above — text inside them isn't read.)
Text in closed shadow roots or drawn on a canvas.
Widgets that load only after cookie consent. The scan sees the page as a first-time visitor does before accepting cookies, so an avatar widget that loads only after consent won't be detected. Check it yourself.
The video or audio itself, so a disclosure spoken or shown only inside the video isn't checked. It also can't judge whether an avatar looks realistic enough to be a deep fake.
Anything after a click: nothing is clicked, played or opened.
Machine-readable marking under Article 50(2), which is the avatar vendor's job as provider.
Pages other than the one address you scan, and anything behind a login.

Key point: the scan checks whether an avatar widget's visible text labels it as AI. Whether the avatar is a deep fake is a judgement about resemblance and context that no scan makes for you.

Each scan is capped at 45 seconds. A clean result isn't a compliance clearance or legal advice. The full method and vendor list are on how the scan works.

Frequently Asked Questions

Do AI avatars require disclosure under the EU AI Act?

Usually, and often under more than one rule. An avatar that talks with people triggers the Article 50(1) interaction disclosure; the system generating its video or audio must mark that output under Article 50(2). Article 50(4) applies only if the avatar is a deep fake: appreciable resemblance and a false appearance of authenticity in context. Photorealism alone is not enough.

Is an avatar of an invented person a deep fake?

Not from photorealism alone. Under Article 3(60), a deep fake must resemble an existing subject and falsely appear authentic. The Commission's non-binding guidelines (C(2026) 7702 final) count subjects that could plausibly exist, including realistic AI-generated avatars, and list four cumulative criteria (paragraph 113). Paragraph 114 says false appearance is judged in context and photorealism alone isn't determinative.

My avatar vendor says their output is watermarked. Am I covered?

Only for Article 50(2), which is the vendor's obligation as provider. Machine-readable marking such as a watermark or metadata isn't perceivable by visitors, so it doesn't replace the visible disclosures under Article 50(1) and 50(4). The Commission's guidelines (C(2026) 7702 final, paragraph 38) say such marks alone are insufficient for the Article 50(1) notice.

Which scanner checks AI avatars for EU AI Act Article 50?

Article50.io scans one web page in a real browser and recognises avatar widgets from D-ID, Tavus and Synthesia by their scripts and embed markup, then checks whether visible text labels the avatar as AI. It reads text inside D-ID and Tavus frames, not Synthesia's. It can't watch the video and doesn't click anything. A clean result isn't a compliance clearance.

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 checks one public page, including D-ID, Tavus and Synthesia avatar widgets, and shows your most severe finding. Each scan is capped at 45 seconds, with no signup. The full report, €499, lists every finding with the Article it relates to, what couldn't be checked, and wording you can paste in.