The Digital Omnibus on AI moved the AI Act's high-risk dates — but it didn't move all of them. This page shows the revised timeline, what changed versus what still lands on 2 August 2026, and a quick check of whether the Article 50 transparency duties apply to you.
The Omnibus deferred the high-risk obligations. It left most of Article 50 — chatbot disclosure, synthetic-content marking, emotion-recognition and deepfake notices — on the original date, and Commission fining powers over GPAI providers begin the same day.
Struck-through dates are the original AI Act dates replaced by the Omnibus. Everything else stands as enacted in 2024.
Regulation (EU) 2024/1689. Staggered application begins.
Prohibited practices (social scoring, manipulation, untargeted scraping…) apply. AI literacy applied here too — the Omnibus later softens its wording (see below).
Technical documentation, copyright policy, training-data summaries for general-purpose AI providers. AI Office and AI Board operational.
Chatbot disclosure, synthetic-content marking for new systems, emotion-recognition and deepfake notices all apply. Commission can start fining GPAI providers. Member State penalty regimes and market surveillance apply. Only exception: Art. 50(2) marking for systems already on the market gets a short grace period.
The Omnibus adds a new prohibited practice to Article 5: AI systems that generate non-consensual intimate imagery or CSAM ("nudifier" apps). Same day, the machine-readable marking duty catches up with generative systems placed on the market before 2 Aug 2026.
Each Member State must have at least one sandbox operational — deadline pushed one year. An EU-level sandbox becomes possible. Sectoral delegated acts for Annex I systems are also due by this date.
The implementing-act template is replaced by Commission guidance, including a template, due by this date.
The full Chapter III regime for stand-alone high-risk systems (employment, credit, biometrics, education, essential services…) now applies from 2 Dec 2027 as a fixed date — no longer conditional on harmonised standards. The Article 27 fundamental rights impact assessment moves with it.
AI as a safety component of regulated products (medical devices, machinery-adjacent…). Note: the Machinery Regulation itself is carved out of the AI Act's scope by the Omnibus.
A deferral, not a dismantling: the risk-based architecture, the obligations and the penalty tiers all survive intact.
The new dates bind only once the Omnibus is published in the Official Journal (entry into force three days later, expected in July 2026). Until then, Regulation (EU) 2024/1689 as enacted remains the law. Use the deferral to build the file — not to pause it.
Five questions against Article 50 as amended. Answers stay on this page — nothing is sent anywhere.
Answer all five questions to see your result.
None of the four Article 50 transparency duties is triggered by your answers. Keep the assessment on file — intended use changes, vendors add generative features, and the answer can flip.
Common exemptions apply (e.g. legally authorised law-enforcement use; assistive or standard-editing functions under 50(2); artistic/satirical works under 50(4); human editorial review for public-interest text). Information must be clear and distinguishable at first interaction or exposure. This is a screening aid, not legal advice.
Verify against the consolidated text on EUR-Lex once the amending regulation is published. This page reflects the state of play as of 11 July 2026.