What Article 50 of the EU AI Act Requires
Article 50 of the EU AI Act creates a direct legal duty to tell people when they are interacting with, or consuming output from, artificial intelligence systems. The duty came into force on 2 August 2026 and applies whenever a natural person is directly exposed to the output of AI systems. This is a transparency obligation separate from the restrictions on prohibited practices under Article 5 and from the broader obligations on high-risk AI systems. Where Article 5 bans certain uses outright, Article 50 instead requires disclosure that allows users to make informed choices about trusting or relying on the output they receive.
The transparency rule covers three main classes of interaction. First, chatbots and conversational AI systems must inform users that they are speaking to a machine, not a human operator. Second, AI systems that generate synthetic media, such as deep fakes, generated photographs or artificial voices, must be marked as synthetic when shared in contexts where a person might reasonably believe the content is authentic. Third, AI systems that infer emotional states or recognise biometric characteristics (such as facial features that signal protected characteristics) must notify the affected individual. The regulation uses precise language about disclosure that must be clear, accessible and timely.
The scope is wide. Article 50 applies to both providers and deployers of AI systems. Providers must build the technical capacity for transparency into their systems at design time. Deployers who use a provider’s system are responsible for implementing disclosure in practice. Where a third party operates the AI system on the deployer’s behalf, responsibility for disclosure sits with the deployer. This means a financial services firm deploying a chatbot through a platform is legally responsible for ensuring the chatbot disclosure is displayed, even if the platform provider built the underlying model.
Enforcement is localised to national authorities in EU member states. The Commission supervises compliance by the model providers themselves. Breaches of Article 50 carry penalties of up to 35 million euros or 7 percent of global turnover, whichever is higher. This is the same fine level as violations of core AI literacy and high-risk system duties, which signals that regulators view transparency as fundamental, not supplementary. The regulation does not specify the exact wording or technical format of disclosure, only that it must meet the test of being clear and easily perceptible.
Disclosure must be timely
Disclosure is only valid if a person is told before or at the moment they first interact with the AI output. Telling someone after the fact that a chatbot conversation was synthetic, or that a video was generated, does not satisfy the legal obligation. This creates practical challenges for deployers who operate mixed teams of human and AI staff. A customer services team cannot ask callers to wait whilst determining whether their query will be answered by a human or a chatbot; the disclosure must happen at the point of connection or conversation initiation.
Scope boundaries
Article 50 does not apply to B2B interactions where both parties are using AI systems with full knowledge. It does not apply to employee training on synthetic content if the employee is explicitly informed of the training context. The rule also does not apply to systems designed to be clearly fictional, such as video game characters or animated avatars, where context makes the artificial nature obvious. Conversely, it does apply to AI systems interacting with children, where the cognitive burden of detecting deception is higher and the duty is therefore stricter.
Interaction versus use
The transparency duty applies when a natural person is directly exposed to the output. This means a researcher using an AI system to analyse data for their own purposes does not require disclosure to themselves. But if that researcher shares the AI-generated analysis with clients or colleagues who rely on it, those recipients must be told the output is AI-generated. An internal audit team running AI-powered controls does not need to disclose to itself; but if the audit findings are reported to the board, the board must know which conclusions came from AI versus human judgement.