EU AI Act transparency duties begin while other deadlines remain later
Europe did not pause the part of the AI Act that hits chatbots and synthetic media. On August 2, 2026, Article 50 transparency duties became generally applicable and enforceable by national authorities, according to a Goodwin alert dated August 3. High-risk system rules once billed for the same day did not land with them. The AI Omnibus (Regulation (EU) 2026/1744), in force since July 27, pushed those obligations out while leaving notices and marking on the original clock.
Vendors who heard “Omnibus delay” and stopped may already be wrong on Article 50. Timeline matches multiple law-firm analyses; confirm edge cases against the Act before hard legal claims.
What Article 50 requires starting Aug 2
Per Goodwin, the duties apply regardless of high-risk classification. Three buckets matter for product teams.
Providers of systems designed to interact directly with people — chatbots, assistants, similar interfaces — must make clear the user is dealing with AI, unless that is obvious from context. Providers and deployers of systems that generate synthetic audio, image, video, or text must mark outputs in a machine-readable, detectable format. Deployers of deepfakes depicting real persons, places, or events face extra disclosure. Deployers of emotion recognition or biometric categorisation systems must inform exposed individuals and handle personal data under EU data protection rules, with limited crime-investigation carve-outs.
If the product talks to users or emits synthetic content in the Union, Article 50 is live. The Commission adopted transparency guidelines on July 20, 2026. Guidance helps implementation; it does not rewrite the statute.
How Omnibus delays change high-risk timing
Standalone high-risk systems in Annex III — employment, education, law enforcement, critical infrastructure, and related uses — now face requirements from December 2, 2027, not August 2, 2026. AI embedded in Annex I product-safety regimes, including medical devices, toys, and lifts, moves to August 2, 2028. Machinery follows a separate path into the machinery regulation by delegated act, due by that 2028 date.
What did not move: prohibitions already in force, GPAI model obligations already running, and Article 50. Goodwin says the Omnibus “does not pause the AI Act’s broader implementation.” Fine mechanics still deserve primary-source checks.
Labelling and deepfake marking stakes
Machine-readable marking is the technical bottleneck. A UI disclaimer alone does not satisfy the synthetic-output duty as the alerts describe it. Provenance tooling and detectable formats sit in the critical path for generative products shipping into the EU.
One narrow grace window exists. Providers of generative systems already on the market before August 2, 2026 get until December 2, 2026 for the marking obligation under the Omnibus. New systems placed on the market from August 2 do not get that cushion. Deepfake disclosure for deployers remains a separate user-facing duty when real people, places, or events are depicted.
Fines under Article 99(4), as cited by Goodwin, can reach €15 million or 3% of worldwide annual turnover, whichever is higher. Enforcement pace in month one is uncertain — medium confidence on pace, higher on the legal start date.
What vendors should ship before audits
Inventory every EU-facing chatbot, generative pipeline, emotion or biometric feature, and deepfake-capable deployer workflow. Wire detectable marking into new generation paths now. Schedule legacy generative stacks against the December 2 marking deadline. Update UX so AI interaction notices are clear unless context truly makes them redundant. Align privacy notices for emotion and biometric tools with GDPR bases.
Use the high-risk extension to build governance, not freeze it. The Omnibus bought time on the hard safety dossier. It did not buy silence on “you are talking to a model” or “this media was generated.”
