Article 50(2) requires a machine-readable, detectable mark. It does not name C2PA. The Commission may treat adherence to the Code of Practice as adequate for 50(2) and 50(4). If you skip the Code, you still need an equivalently adequate technical means.

Do we label every autocomplete token?

Assistive standard-editing that does not substantially alter the input or its semantics is outside 50(2). A three-word grammar fix is not a generative system output in that sense. A rewritten press release is.

AI-generated content labeling requirements (EU Article 50)

Labeling under Article 50 is two jobs. Providers mark synthetic audio, image, video, or text so machines can detect it (50(2)). Deployers who publish deepfakes or public-interest text tell humans (50(4)). Mixing those up is how teams ship a watermark and forget the visible disclosure — or the reverse.

Educational resource only — not legal advice. scm.cc publishes a readiness report and engineering gap analysis for AI features. It does not give legal advice, represent you before a regulator, or state that a product is ready to ship under any law. This page is educational product guidance, not legal advice. It is not a readiness report, not an engineering gap analysis of your systems, and not a determination that any product meets a legal duty. Confirm current official text before you ship.

Last checked against official sources: 2026-10-11.

Provider marking (50(2))

Outputs must be marked in a machine-readable format and detectable as artificially generated or manipulated. The technical solution must be effective, interoperable, robust and reliable as far as technically feasible, taking into account content type, cost, and the state of the art. Standard editing that does not substantially alter the input is out.

Systems already on the market before 2 August 2026 have until 2 December 2026 (Article 111(4)). New generative systems placed on the market from 2 August 2026 needed the mark from day one.

Deployer labeling (50(4))

If you generate or manipulate image, audio, or video that constitutes a deepfake, disclose that it was artificially generated or manipulated. Artistic, creative, satirical, fictional or analogous works get a lighter “this work contains generated or manipulated content” disclosure that does not hamper enjoyment.

If you publish text to inform the public on matters of public interest, disclose that the text was artificially generated or manipulated, unless a human review or editorial-control process ran and a natural or legal person holds editorial responsibility.

What to put in the product

A visible human-readable label at first exposure, plus a machine-readable hint (C2PA / Content Credentials, an alt-text convention, or a file-sidecar — pick what your stack can actually persist). The labeling snippet page has copy and markup. The mark is a provider engineering task; the visible label is often a deployer publishing task. Many SaaS products are both.

Questions teams ask

Is C2PA mandatory?: Article 50(2) requires a machine-readable, detectable mark. It does not name C2PA. The Commission may treat adherence to the Code of Practice as adequate for 50(2) and 50(4). If you skip the Code, you still need an equivalently adequate technical means.
Do we label every autocomplete token?: Assistive standard-editing that does not substantially alter the input or its semantics is outside 50(2). A three-word grammar fix is not a generative system output in that sense. A rewritten press release is.

Back to the AI law hub

Primary sources

Official text wins. Last checked 2026-10-11. This page is educational product guidance, not legal advice. It is not a readiness report, not an engineering gap analysis of your systems, and not a determination that any product meets a legal duty. Confirm current official text before you ship.