AI Act Article 50(1) and a rendered avatar video: the test

Article 50(1) is a test about interaction, and a finished avatar video is a rendered file, not a conversation. The Commission FAQ sets four cumulative criteria, and one of them is a system "designed for a genuine two-way exchange with people". Sume's avatar video endpoint takes a script and returns a video, so it sits on the one-way side of that criterion. This is a reading aid, not legal advice.

The criteria are from the Commission's Article 50 FAQ; the Sume facts are from Generate avatar video, both read 2026-10-01.

What are the four criteria in Article 50(1)?

The FAQ says the obligation to inform people applies when all four hold. Failing any one means 50(1) does not apply to that system.

Is a script-driven avatar video a two-way exchange?

The docs open with: "Avatar videos turn a ready avatar into a script-driven talking video." You send avatar_handle and exactly one of script or video_inputs; Sume estimates the duration (4-60 seconds inclusive) and renders one MP4. Nothing in that request lets a viewer reply to the avatar. That makes the second criterion the one to examine, and it is a question about how you deploy the file.

What changes if I put the avatar in a live chat?

If your own product wires an avatar to live replies, the product you build is a different system, and you should assess that system against the same four criteria. The FAQ also says people need not be told when it is obvious they are interacting with AI, and that this exception is read restrictively.

Creating the avatar from a prompt (POST /v1/avatar-1.0/generate) does not change this; the docs list Prompt, Profile and Image as ways to make the avatar, and each creates a job.

Does passing 50(1) mean I never label the video?

No. Article 50(1) is only the interaction rule. Whether a rendered video needs a label is a separate question under the deepfake and marking rules; see what creators must do and the three deepfake criteria.