AI presenter videos (the Synthesia/HeyGen genre of a realistic human avatar delivering your script) are one of the cleaner cases under Article 50(4): a real-seeming person who doesn't exist, or wasn't actually filmed saying this, is exactly the content the deepfake duty describes. The duty is triggered by the video falsely appearing authentic, not by any intent to deceive.
That includes the flattering case people hope is exempt: an AI clone of your own founder, made with full consent, delivering a message they approved. Consent settles the personality-rights question; it doesn't touch the AI Act question, which is about the audience knowing the footage is synthetic. If viewers would take it for a real recording, it needs the disclosure.
The format, per the June 2026 Code of Practice: an opening disclaimer plus a persistent "AI" icon ("KI" German, "IA" French) for the duration of the video. Front-loading a one-second flash of text does not satisfy the persistent-icon half.
What's out: obviously stylised or cartoon avatars nobody would mistake for footage of a person. The test is the reasonable viewer's perception of authenticity, not the tool used. A deliberately synthetic-looking assistant character is a different creature from a photoreal presenter.
If the avatar video is an ad, note that the artistic carve-out won't help (it excludes primarily commercial content), and if the avatar also talks back to users interactively, the Article 50(1) chatbot disclosure duty stacks on top. Keep a copy of the labelled master and a dated note of placements as evidence.
Not sure how this applies to your specific setup? The free 2-minute check tells you which of the four Article 50 duty areas likely apply to your company. No email required.
Take the free check Generate disclosure text