AI Transparency Obligations Under Article 50

Article 50 of the EU AI Act imposes transparency obligations on providers and deployers of certain AI systems that interact with or affect natural persons. These obligations are distinct from — and in addition to — the high-risk AI obligations under Chapter III. They apply to limited-risk AI systems that fall outside Annex III but still interact with people in ways that require disclosure.

The four main transparency obligations under Article 50 cover: AI systems that interact directly with natural persons (chatbots, virtual assistants); AI systems that generate synthetic content (text, images, audio, video); AI systems that perform emotion recognition or biometric categorisation; and AI systems that generate or manipulate deepfake content.

Providers of AI systems intended to interact directly with natural persons must ensure those systems are designed so that natural persons are informed they are interacting with an AI system, unless this is obvious from the context. This obligation applies at the beginning of the interaction — not buried in documentation or disclosed only if asked.

The obligation applies to all AI systems designed for human interaction: customer service chatbots, AI-powered virtual assistants, automated telephone response systems, and any other system that converses with users in natural language.

Providers of AI systems that generate synthetic text, images, audio, or video content must ensure those outputs are machine-readable labelled as AI-generated. This applies to text generation systems, image generation models, audio synthesis, and video generation tools — including deepfake generation.

The labelling obligation uses technical watermarking or metadata standards. The European AI Office is developing harmonised technical standards for AI-generated content labelling. Until these are finalised, providers should implement available watermarking solutions and document their implementation approach.

AI systems that infer emotions or categorise people based on biometric data face some of the strictest transparency obligations under the Act. Deployers of such systems must inform persons who are exposed to them. Deployers must similarly inform the persons being processed.

Emotion recognition systems are broadly prohibited in workplaces and educational institutions, with the only exception being narrow medical or safety reasons (Article 5(1)(f)). Biometric categorisation systems that categorise people based on protected characteristics — including political opinions, religious beliefs, or sexual orientation — are prohibited entirely under Article 5.