Brussels scopes its deepfake marking duty by actor, not by tool

The shot

A synthetic or manipulated video circulates in the European Union after 2 August 2026, the date the European Commission's own transparency guidance says Article 50 of the EU AI Act starts to apply. The provision does not ban the video; it requires certain actors to mark or disclose it. The Commission's guidelines on AI transparency obligations and its own FAQ on Article 50 are the regulator's account of who has to do what, as retrieved on 16 September 2026.

What the documents show

The guidance separates two roles. Providers of AI systems that generate synthetic audio, image, video or text must ensure outputs carry a machine-readable mark detectable as artificially generated or manipulated. Deployers, the organizations that put a system to use, carry a separate duty: when they use a system to create or manipulate content that qualifies as a deepfake, defined as AI-generated or altered audio, video or image resembling real people or events that would falsely appear authentic to a person, they must disclose that the content is artificially generated. Marking compliance for systems already on the market before 2 August 2026 has a grace period running to 2 December 2026, per the FAQ, which also lists exemptions scoped narrowly: standard editing functions, limited business-to-business contexts, and deepfakes in evidently artistic, satirical or fictional works. This is the regulator's own reading of its statute, not a court ruling on any specific video, and it does not mean every generative video use in the EU now carries a disclosure duty regardless of context.

The workflow

For a production or platform operating under the Act, the guidance implies a two-part check: first, is the tool a provider-side generation system that must embed a machine-readable mark at output, and second, is the finished piece being deployed in a way that meets the deepfake definition, triggering a separate deployer disclosure regardless of the underlying mark. A studio using a text-to-video model for a clearly fictional sequence would consult the artistic-work exemption named in the FAQ; a studio depicting a real person doing or saying something they did not would not qualify for that exemption on the guidance's own terms.

What the tool does not change

Enforcement, per the guidance, sits with national market surveillance authorities, the AI Office and the European Data Protection Supervisor, not with the marking technology itself. Whether a given production actually qualifies for an exemption, and whether a mark is applied correctly, remains a judgment call by producers and regulators that the machine-readable mark cannot make on its own.

• Does this production's use of generative video fall under the provider marking duty, the deployer disclosure duty, or both?
• Would this specific depiction plausibly meet the guidance's own deepfake definition, or does an exemption apply?
• Who in the pipeline is responsible for confirming the machine-readable mark actually survives to publication?

Article 50 gives European regulators a transparency floor with named exceptions, described here as the Commission's own guidance states it on 16 September 2026, not as an editorial prediction of how any single case will be decided.