Article 50 applies from August 2, 2026.
Providers and deployers carry four distinct transparency duties, subject to express exceptions. The adopted Digital Omnibus text awaits Official Journal publication and entry into force. Once effective, its targeted transition runs to December 2, 2026 only for Article 50(2)’s marking and detection duty on qualifying generative systems placed on the market before August 2. Here is the durable rule map and the boundary between compliance and supporting evidence.
Four duties that are not limited to high-risk systems.
Article 50 reaches some ordinary products as well as high-risk systems. Article 50(1) covers systems intended to interact directly with people; Article 50(2) separately covers providers of systems generating specified synthetic content; Articles 50(3)–(4) assign deployer duties. Article 50 generally applies from August 2, 2026. The adopted Digital Omnibus text awaits Official Journal publication and entry into force. Once effective, it adds a targeted transition to December 2, 2026 for Article 50(2)’s marking and detection duty on qualifying systems placed on the market before August 2.
Four obligations, two roles.
A factual paraphrase; the regulation’s text governs. Provider = the actor that develops or has the system developed and places it on the market or puts it into service under its own name or trademark. Deployer = the actor using it under its authority, excluding personal non-professional use.
| Clause | Who | Obligation |
|---|---|---|
| Art. 50(1) | Providers | Systems intended to interact directly with people must be designed so those people are informed they are interacting with AI, unless that is obvious to a reasonably well-informed, observant, and circumspect person in context. A qualified law-enforcement exception also applies. |
| Art. 50(2) | Providers | Providers of AI systems, including general-purpose AI systems, that generate covered synthetic audio, image, video, or text must make the outputs machine-readably marked and detectable through effective, interoperable, robust, and reliable solutions as far as technically feasible. Editing, non-substantial-alteration, and qualified law-enforcement exceptions apply. |
| Art. 50(3) | Deployers | Deployers of emotion-recognition or biometric-categorisation systems must inform exposed people and process personal data under applicable Union law. A qualified law-enforcement exception applies. |
| Art. 50(4) | Deployers | Deployers must disclose covered deepfakes. Covered public-interest text must be disclosed unless it underwent human review or editorial control and a natural or legal person holds editorial responsibility. Artistic-content and qualified law-enforcement regimes also apply. |
Article 50 information must be clear and distinguishable, supplied no later than the first interaction or exposure, and conform to applicable accessibility requirements (Art. 50(5)). Penalties can reach EUR 15 million or 3% of worldwide annual turnover, whichever is higher for an undertaking (Art. 99(4)); the SME cap uses the lower applicable amount (Art. 99(6)).
Three surfaces, one record.
1. The disclosure surface
Where Article 50(1) applies, provide clear and distinguishable information no later than the first interaction, unless the AI nature of the interaction is already obvious in context. The regulation does not prescribe one UI affordance.
2. Machine-readable marking
Covered outputs must be machine-readably marked and detectable through an effective, interoperable, robust, and reliable technical solution. The final guidance permits compliant marking post hoc, at the model, or during inference.
3. The record that it happened
A dated, tamper-evident record can preserve evidence that a disclosure event was recorded. It is supporting evidence, not a substitute for the disclosure, timing, accessibility, marking, and detection requirements.
The honesty boundary, as everywhere on this site: Caisson generates technical evidence; whether your surfaces satisfy Article 50 is a separate legal determination that stays with you and your counsel.
How the AI Act phases in.
- August 1, 2024: Regulation (EU) 2024/1689 enters into force.
- February 2, 2025: prohibited-practice bans apply.
- August 2, 2025: general-purpose AI model obligations apply, except Article 101.
- August 2, 2026: the general application date. Article 50 generally applies. The other Article 50 duties were not postponed.
- December 2, 2026: under the adopted Digital Omnibus text, once it enters into force, providers of generative AI systems placed on the market before August 2 must conform with Article 50(2)’s marking and detection duty.
Article 50(2) outputs and Article 50(4) deepfakes generated or manipulated before August 2, 2026 do not require retroactive marking or labelling. Public-interest text follows that rule only when it was both generated or manipulated and published before the cutoff.
What the July 2026 final guidance settled.
The writing piece isolates the August-versus-December question, the mixed-system example, and the separate rule for pre-existing content. This page remains the evergreen reference for the duties, roles, exceptions, and implementation boundary.
Read the dated guidance analysisPrimary sources.
Reviewed 2026-07-27. The final Commission guidelines are non-binding; only the Court of Justice of the European Union can ultimately give an authoritative interpretation of the AI Act. The adopted Digital Omnibus text awaits Official Journal publication and entry into force.
- Regulation (EU) 2024/1689 on EUR-Lex — Articles 3(3)–(4), 50(1)–(5), 99(4), 99(6), 100(3), and 113; OJ pp. 82–83, 115–117, and 123/144
- European Commission final-guidelines landing page — Page header “Publication 20 July 2026”; introductory paragraph beginning “The Commission adopted these guidelines”
- European Commission final Article 50 guidelines — Paragraphs (5), (6), (69)–(74), (151), and (153)–(154), pp. 3–4, 24–25, and 49–50
- European Commission quick facts on AI transparency — “Enforcement and penalties” → “Surveillance authorities”, “Penalties”, and “Exceptions”
- Council adopted Digital Omnibus legislative text — Article 111(4), p. 90/102, and Article 4, p. 101/102: systems “placed on the market before 2 August 2026”; compliance with Article 50(2) by 2 December 2026; entry into force on the third day after Official Journal publication
- Council final-approval and next-steps notice — Page header dated 29 June 2026; opening paragraph beginning “Today, the Council gave its final green light”; “Next steps” paragraph on Official Journal publication and entry into force
Common questions.
Does Article 50 apply to my SaaS chatbot or AI agent?
Was the August 2, 2026 date delayed?
What are the penalties for non-compliance?
Does Caisson make my product Article 50 compliant?
Do pre-existing outputs need retroactive labels?
The audit chain, evidence bundles, and versioned configuration that preserve evidence of Article 50 disclosure events ship in the Compliance bundle, wired and testable from day one. That evidence supports review; it does not determine legal satisfaction.