EU AI Act's content-labelling rules take effect today

From 2 August the AI Act's Article 50 transparency duties bite across the EU: machine-readable marking, chatbot disclosure and deepfake labels.

What became binding

The European Union's AI Act crossed its most consequential milestone for generative-AI products on 2 August 2026, when the transparency obligations in Article 50 became directly applicable across all 27 member states.

Providers of generative systems must now mark their outputs in a machine-readable way so that synthetic images, audio, video and text can be detected downstream. Systems that interact directly with people — chatbots, voice assistants, agent front ends — must make clear that the counterpart is a machine, unless that is obvious from context. On the deployer side, anyone publishing a deepfake must disclose that the material is artificially generated or manipulated, and AI-generated text published on matters of public interest must be labelled where it has not passed human editorial review. Deployers of emotion-recognition and biometric-categorisation systems must inform the people exposed to them.

Two timing carve-outs are worth keeping straight. The rules are not retroactive: content produced before 2 August does not have to be relabelled after the fact. Separately, a limited grace period applies to the marking duty itself — generative systems already placed on the EU market before 2 August have until 2 December 2026 to comply with machine-readable marking. Every other Article 50 duty, including chatbot disclosure and deepfake labelling, applies from today.

Enforcement and the missing half

Enforcement falls mainly to national market surveillance authorities rather than the Commission's AI Office, whose remit is limited to systems built on general-purpose models by the same provider, or systems folded into very large online platforms and search engines. The European Data Protection Supervisor covers EU institutions. Infringements of Article 50 can draw fines of up to €15 million or 3 percent of worldwide annual turnover, whichever is higher, with proportionality allowances for SMEs and small mid-caps.

The date arrives with a large hole in it. Under the Digital Omnibus on AI — provisionally agreed on 6 May 2026 and confirmed by member-state representatives in the Council a week later — obligations for stand-alone high-risk systems listed in Annex III slip to 2 December 2027, and for AI embedded in regulated products under Annex I to 2 August 2028. What was meant to be the AI Act's main compliance wave has become, in practice, a transparency-only wave.

Companies can lean on the Commission's voluntary Code of Practice on Transparency of AI-generated Content, published on 10 June 2026, which the Commission and AI Board have endorsed as an adequate compliance route. The Commission says about 190 organisations had signed by the end of July 2026 — but that headline number covers two separate commitments rather than a single roster, and the two sections are signed independently of each other.

Section 1, with 82 signatories, is the provider-side undertaking, covering the Article 50(2) duties on machine-readable marking and detection. The Commission's published examples for it are Aleph Alpha, Anthropic, Black Forest Labs, Cohere, Google, Meta, Microsoft, Mistral, OpenAI and Synthesia. Section 2, with 152 signatories, is the deployer-side undertaking, covering the Article 50(4) duties on disclosing deepfakes and labelling AI-generated text published to inform the public on matters of public interest. Its published examples are Bulgari, Deutsche Lufthansa, Fastweb, Getty Images, Iberdrola and Lenovo. The Commission's summary lists the two sets separately and does not name Google, Meta or Microsoft among the Section 2 examples, so their deployer-side status should not be inferred from the provider list; the full roster is the only place to confirm which organisations signed both. The Commission notes that roughly half of all signatories are small or recently founded companies.

The code stays open for signature indefinitely — the 27 July cut-off applied only to inclusion in the initial published list, and the Commission says the published roster will keep being updated — so both counts should be read as a snapshot rather than a closed list. Signing remains voluntary either way, and the Article 50 duties bind signatories and non-signatories alike.

Why it matters

This is the first time a major jurisdiction has made provenance marking a legal duty for frontier generative models rather than a voluntary pledge. Because watermarking and disclosure are baked into model serving rather than regional UI, the practical effect will be global: EU rules now shape what every user of a large image, video or voice model sees.

← All briefings