AI Act deepfake definition: the three cumulative criteria

Under the Commission's Article 50 FAQ, AI-generated or manipulated image, audio or video is a "deepfake" (Article 3(60)) only when three criteria are all met: resemblance to the subject, an existing (or plausibly existing) person, object, place, entity or event, and a false appearance of being authentic or truthful. Miss any one and the content is not a deepfake under that definition.

This post walks an AI avatar video through the three tests. It is a reading of the FAQ as published, not legal advice. Wording is from the Commission page, read 2026-10-01; Sume behavior is from Generate avatar video and Face swap (Beta).

What are the three criteria exactly?

The FAQ lists them as cumulative. The table keeps the Commission's wording short.

How is the third criterion assessed?

The FAQ says the Guidelines let you weigh the level of resemblance, the potential substantive message, the intended and foreseeable deployment contexts, and the intended audience and what it expects. If the audience in that context does not expect the content to be authentic, it may not falsely appear authentic. The FAQ's example is background scenes, special effects or technical pre- and post-processing in standard movie production.

Where does a script-driven avatar video land?

Sume's docs describe avatar videos as turning "a ready avatar into a script-driven talking video." Whether that output meets the criteria depends on your case. An obviously stylized or cartoon presenter may fail the resemblance or false-appearance test, but the FAQ's existing criterion also covers subjects that could plausibly have existed, so a photorealistic invented person is not automatically outside it. An avatar built to look like a real, identifiable person that viewers would take as filmed footage is the situation the three tests are aimed at.

Face swap changes the picture. Avatar Face Swap is a Beta endpoint that applies a ready avatar face onto a public source video, so the output carries a real source clip's scenes and motion. That makes the resemblance and existing-subject questions worth asking deliberately. For the label side of that case, see eu-ai-icon partially AI-modified vs fully AI-generated.

What should I do with the answer?

Write down, per video, which of the three criteria you judged and why. If all three are met, the FAQ says deployers must disclose at first exposure in a clear, perceivable way; see first exposure and machine-readable marking. If you are unsure, the FAQ's own pointer is the Guidelines on Transparency of AI-Generated Content. For a legal conclusion on a specific campaign, ask counsel.