AI Act Article 50(2): 2 December 2026 grace period, in plain terms

Article 50(2) is the provider duty to mark AI outputs in a machine-readable way. The Commission's FAQ says a limited grace period is envisaged only for AI systems placed on the market before 2 August 2026, and only for that marking and detection duty; those providers must comply from 2 December 2026. It also says content generated before 2 August 2026 does not need to be labelled retroactively. This is not legal advice.

Dates are from two Commission FAQ pages; what Sume records is from its docs and OpenAPI. All read 2026-09-30.

What are the dates?

Who does the 2 December date apply to?

Providers of those systems, for marking and detection under Article 50(2). The FAQ does not extend it to a deployer's disclosure of a deepfake, which it says must happen at first exposure at the latest. If you publish AI video, treat 2 August as your date and ask a lawyer whether the later one touches you at all.

How do I show when a file was made?

Sume's docs do not describe a mark inside outputs, so use the records it does keep. The usage ledger rows in the OpenAPI document are scoped to the authenticated workspace and carry job_id and created_at. A job also has events: GET /v1/jobs/:id/events is described as a public timeline for debugging and recovery, starting with job.created.

Store the job id and the Sume URL next to each delivered file, as in what to log per generated file. See Jobs and results for the event list.

Does a pre-August file need a label?

Per the FAQ, not retroactively, though the Commission encourages it. A file you edit or republish after 2 August is a separate question, so keep the original job record and the date you changed it. See also the Article 50 labeling guide for creators.