AI Act Article 50(3): what Sume video inspect returns
Video inspect returns probe facts, stills and optional STT. The FAQ ties Article 50(3) to emotion recognition; the inspect docs list no such output.
Sume's video inspect returns probe facts, sampled stills and optional speech-to-text for one clip; its docs do not list emotion or biometric categorisation as an output. Article 50(3) of the AI Act, as the Commission FAQ explains it, applies to deployers of emotion recognition and biometric categorisation systems. This post is a plain reading of public pages, not legal advice.
Scope text is from the Commission's Article 50 FAQ; Sume facts are from Video inspect, all read 2026-10-01.
What does Article 50(3) cover?
The FAQ says deployers of emotion recognition systems and biometric categorisation systems must inform the natural persons exposed to those systems of their operation, to protect their privacy. It adds that this does not mean explaining, for example, the system's purpose, and that the obligation applies whether people are exposed in real time or the system runs after the fact.
What does video inspect return?
Video inspect reads one media.sume.com clip already owned by the workspace. Its page says inspect returns probe facts, stills and optional STT, and that it is not typed scenes: semantic questions are separate tools that appear only when listed in tools_list. It never re-encodes the source and never produces an MP4.
The page documents no emotion, expression or identity field in the result.
| Tool | Output per the docs |
|---|---|
POST /v1/video-inspect | Probe facts, stills, optional STT; probe and stills unbilled |
POST /v1/video-frames | Durable stills at times you name (Video frames) |
POST /v1/audio-detach | A new audio artifact from one workspace video; the video is untouched |
Does using inspect make me an Article 50(3) deployer?
The cited pages do not say, and a tool name does not settle it. Article 50(3) is about systems that recognise emotions or categorise people biometrically. If you build a feature on top of inspect output that does that, you own that design and should take advice.
If you only read duration, resolution and a transcript to prepare edits, you are using the documented outputs listed above.
What should I check in my own pipeline?
List every field you store from an inspect or frames call, and every model you run on the stills. If any step infers feelings or traits of a person, treat it as a separate system and check the FAQ text. Relevant context on preview versus final files is in closed-loop previs and the final output line.
Sources
Related posts
More in Developers
- EU AI Content Code: Section 1 providers or Section 2 deployers?
Section 1 of the EU Code is for providers marking AI output; Section 2 is for deployers labelling deepfakes. A team shipping API video starts with its own role.
- AI music exactly 30 seconds: no duration field, use Timeline
Music Router rejects duration and duration_seconds. Ask for the length in the prompt, then fix the exact output length with Timeline 1.0 audio.duration_seconds.
- Nova Reel start_async_invoke S3 output vs a Sume result URL
Nova Reel's start_async_invoke writes output.mp4 to your S3 bucket and needs IAM. A Sume job returns a result_url and durable media.sume.com links.
- Avatar video captions error over 60 seconds: split the script
Inline captions on an avatar video are rejected when the estimated duration is over 60 seconds, the same cap as the job. Split long scripts into jobs.
Written by Sume