The debate around the EU AI Act's Digital Omnibus has created a widespread misimpression: that compliance deadlines have been pushed back across the board. They have not. The Omnibus provisionally defers certain obligations for Annex III high-risk AI systems. Article 50 — the transparency and marking obligations for generative AI and AI systems that interact with people — was not in scope for deferral. It applies from 2 August 2026 regardless of whether or when the Omnibus is formally adopted.
What Article 50 requires
Article 50 creates four categories of obligation:
1. AI systems interacting with natural persons. Providers of AI systems designed to interact with humans (including chatbots and voice assistants) must ensure those humans are informed they are interacting with AI, at the time of first interaction. Exception: where this is obvious from context or where the system is used for the prevention, detection, investigation, or prosecution of criminal offences.
2. Emotion recognition and biometric categorisation. Deployers of emotion recognition systems or AI systems that categorise individuals on the basis of biometric data must inform the persons exposed to such systems. This applies subject to the Article 5 prohibition, which bans certain biometric categorisation systems entirely.
3. Synthetic content — images, audio, video. Providers and deployers of AI systems that generate or manipulate synthetic audio, image, or video content (including deepfakes) must mark and make detectable as artificially generated or manipulated content that: portrays real people; shows real places or events; could deceive a reasonable person. Machine-readable marking is required. An explicit human-readable label is required for content that constitutes a deepfake.
4. Text on matters of public interest. Providers and deployers of AI systems that generate text published for the purpose of informing the public on matters of public interest must disclose that the text is AI-generated. This applies unless the text has undergone substantial human review and a natural person is editorially responsible for it.
The Code of Practice on marking — published 10 June 2026
The European Commission published the final Code of Practice on marking and labelling AI-generated content on 10 June 2026, ahead of the August deadline. The Code is voluntary: signing is not required by law. But non-signatories will face higher scrutiny from enforcement authorities, who will conduct case-by-case gap analyses rather than treating compliance with the Code as presumptive compliance with Article 50.
The Code establishes technical requirements for machine-readable marking: watermarking embedded in content, metadata standards including C2PA (Coalition for Content Provenance and Authenticity), logging facilities and fingerprinting. It covers both upstream marking (by providers, embedded at generation) and downstream labelling (by deployers, visible to audiences). Open-weight model providers are addressed: despite their exclusion from some AI Act provisions, the Code encourages structural marking encoded during training to enable downstream compliance.
What organisations need to have done by 2 August 2026
For organisations that provide or deploy AI systems within the EU (or accessible to EU users), the practical requirements are: identify every AI system that interacts with natural persons and confirm disclosure mechanisms are in place; identify every generative AI system producing synthetic content and confirm machine-readable marking is implemented; map deepfake-capable use cases and confirm human-readable labels are in place; and review third-party AI tool agreements to confirm vendors are providing marked outputs. Organisations that have not yet assessed their Article 50 exposure should treat the remaining weeks as the implementation window.
The enforcement picture is split: the European AI Office oversees general-purpose AI model providers; national market surveillance authorities in each member state oversee deployers. By 1 June 2026, the AI Office launched its enforcement function with independent expert support — the first signal that the Office is moving from setup to active supervision.
Primary sources: European Commission — AI Act (official page, updated June 2026) | Commission — Code of Practice FAQ | EU AI Act — Article 50 Transparency Rules Guide