The EU's AI Video Labelling Rules Are Live: What Article 50 Asks of You

If you publish AI-generated video that reaches viewers in the European Union, you are now the party responsible for telling them it is AI-generated. Article 50 of the EU AI Act has applied since August 2, 2026, and the Commission's reading of it covers individual professional creators, not only platforms, as Euronews reported when the rules came into force.

What Article 50 Requires, in Plain Terms

The article splits the duty in two. Providers of generative systems, the companies building the models, must mark their output in a machine-readable format so it can be detected as artificially generated. Deployers, the people and companies using those systems, must disclose when content is a deepfake or has been manipulated. The Commission's guidelines set out which paragraph binds whom.

Timing is part of the obligation. Disclosure has to be clear, easy to understand and perceivable from the outset, at the latest before someone is exposed to the content, under Article 50(5).

Who Has to Label What

A brand, an agency, a publisher or a commercially acting creator disclosing a synthetic image or video falls inside the deployer definition. So does a freelance video maker whose work is published professionally, regardless of where the maker is based, because the test is whether the content reaches the EU market.

Four carve-outs matter. Purely personal, non-professional use is outside scope. Evidently artistic, creative, satirical or fictional work carries a lighter duty: the disclosure must be made in a way that does not hamper the enjoyment of the work. Text that has been through genuine human editorial review by a named person is exempt from the text-labelling duty. Law-enforcement use authorised by law is also outside the Article 50 duties.

What a Label Has to Look Like

The Commission published a set of icons: a basic AI mark, one reading Fully AI-Generated, and one reading Partially AI-Modified, each in several colour variants. Using them is optional. Disclosing is not, and the Commission's own guidance is that icon use alone does not establish compliance. A valid disclosure is one a reasonable viewer notices, stays with the content when it is shared, and does not sit behind an overlay.

The Deadline Most People Missed

August 2 was the application date for new systems. The regulation gives systems already on the market before that date a transition for the machine-readable marking duty, and it runs to December 2, 2026, as the compliance analysis at Securing.AI sets out. If you generated a library of synthetic video before August on a model that does not embed provenance data, that date is the one on your calendar.

Alongside the law sits the voluntary Code of Practice on Transparency of AI-Generated Content, published on June 10, 2026. Roughly 190 organisations had signed it by the end of July, a list that includes Google, Meta, Microsoft, Mistral, OpenAI, Anthropic, Cohere and Synthesia. Signing creates a presumption of compliance rather than a replacement for it, and it commits signatories to a layered approach: signed provenance metadata plus imperceptible watermarking, with a watermark-detection interoperability solution due by February 2, 2027. The code's watermarking commitment applies to free-form text above roughly 200 tokens, which exempts very short passages.

Why the Watermarks Underneath Are Fragile

The same analysis is blunt about the mechanism. Metadata is stripped routinely when a file is re-saved, screenshotted or re-encoded, and imperceptible image watermarks can be removed by passing the image through a generative model again, without needing anyone's key. Anthropic makes a similar admission about its own text marking on its support page for how Claude marks AI-generated content, saying extensive rewriting or paraphrasing may reduce or remove the detectable signal.

That is the practical lesson for a creator. Article 50 is a labelling obligation, not a detection capability. The visible disclosure is the part that survives a re-upload, which is the part a viewer actually sees.

What It Means If You Publish Into the EU

Enforcement sits with national market surveillance authorities, with the AI Office for the systems it supervises, and with the European Data Protection Supervisor where an EU institution is the provider or deployer. The ceiling on a breach is 15 million euros or 3% of total worldwide annual turnover, whichever is higher, and small and medium-sized firms get the lower of the two figures, under Article 99 of the regulation.

The forward look is the February 2, 2027 commitment, when signatories of the code are meant to offer interoperable detection, and the December 2, 2026 transition for older systems. Until both land, the disclosure in the picture is doing the work the watermark cannot.

The Commission's icons and the full article text are at artificialintelligenceact.eu; the voluntary code's signatory list is published by the Commission.