What it is
Article 50 of the European Union’s Artificial Intelligence Act is a set of disclosure rules, not a risk-classification regime. It does not ask whether an AI system is dangerous; it asks whether the people encountering that system know they are dealing with one. The obligations apply in four situations: AI systems that interact directly with people, such as chatbots, must disclose that they are AI; providers of systems that generate synthetic audio, image, video, or text must mark that output in a machine-readable format so it can be detected as artificially generated; deployers using AI for emotion recognition or biometric categorisation must tell the people being analysed; and deployers who publish AI-generated deepfakes, or AI-generated text on matters of public interest, must disclose that origin.
The rules fall on two different parties depending on the obligation. Providers — the companies that build the system — carry the technical burden of embedding disclosure and machine-readable marking into the product itself. Deployers — the companies and institutions that put the system to use — carry the burden of telling the humans in front of them what they are looking at. This split matters editorially: a chatbot vendor and the bank that deploys that chatbot can each be in or out of compliance independently of the other.
Article 50 became legally binding on 2 August 2026, two years after the AI Act’s August 2024 entry into force. A narrower grace period applies only to the machine-readable marking requirement for generative systems already on the market before that date: those providers have until 2 December 2026 to implement watermarking. The disclosure obligations themselves — telling a user they are talking to a machine — carried no such grace period and applied immediately on 2 August.
Why it matters for AI governance and narratives
Article 50 is a useful marker in the AI Act’s own framing contest because it is the provision the EU chose to protect when it otherwise retreated. In November 2025 the European Commission proposed a “Digital Omnibus” package — now Regulation (EU) 2026/1744, in force since 27 July 2026 — that deferred the Act’s toughest tier, the high-risk system obligations (risk management, technical documentation, human oversight, conformity assessment for uses like biometric identification, employment screening, and law enforcement), from August 2026 to December 2027, with product-embedded high-risk systems pushed to August 2028. That deferral was publicly framed by Brussels as a response to industry complaints about implementation readiness and a lack of finalised harmonised standards.
The Omnibus did not touch Article 50. That selective protection is itself a signal of where the Commission judged its political and legal exposure to be lowest: transparency obligations are comparatively cheap for industry to implement and popular with the public, while the high-risk compliance architecture is the part large AI developers lobbied hardest against. For an observatory tracking how the meaning of “AI regulation” is contested, the Omnibus split offers a clean test of whose framing prevailed — industry’s argument that the Act moved too fast won on substance (deferral of the costly parts) while regulators kept the symbolically visible disclosure requirement intact.
Key facts and dates
- The AI Act entered into force in August 2024; Article 50’s transparency obligations became applicable on 2 August 2026.
- A grace period for machine-readable marking of generative AI systems already on the market runs until 2 December 2026; the underlying disclosure duty had no such delay.
- The Digital Omnibus, Regulation (EU) 2026/1744, entered into force 27 July 2026 and deferred high-risk AI obligations: stand-alone high-risk systems (Annex III, covering biometric identification, critical infrastructure, employment, and law enforcement uses) move to 2 December 2027; high-risk systems embedded in regulated products (Annex I) move to 2 August 2028.
- Non-compliance with Article 50 can draw administrative fines of up to €15 million or 3% of worldwide annual turnover, whichever is higher.
- In France, enforcement of the transparency duties is divided among three bodies: the data protection authority CNIL, the consumer-protection authority DGCCRF, and the broadcasting regulator Arcom — illustrating how a single EU-wide obligation fragments into different national enforcement architectures as member states implement it.
Where to learn more
- European Commission: AI Act regulatory framework — the Commission’s own policy page on the Act.
- European Commission: guidelines on transparency obligations — official guidance on Article 50 specifically.
- EU AI Act Explorer: Article 50 practical guide — detailed breakdown of the four obligation categories and timelines, maintained by the Future of Life Institute’s AI Act tracking project.
- Gibson Dunn: EU AI Act Omnibus Agreement — law firm analysis of what the Digital Omnibus changed and left unchanged.