EU AI Act Transparency Requirements

Article

50 distinguishes four categories: First, interactive AI systems like chatbots must disclose that users are communicating with AI (Art. 50(1)). Second, providers of AI systems generating synthetic content (images, audio, video) must label outputs as AI-generated in a machine-readable format (Art. 50(2)). Third, operators must label deepfakes as manipulated content (Art. 50(4)). Fourth, AI-generated text on matters of public interest must be disclosed. These obligations apply from August 2, 2026.

Article

13 applies to high-risk AI systems and requires comprehensive technical documentation, clear instructions for use, and information about performance characteristics and limitations. Article

50 applies to limited-risk AI systems and primarily governs labeling: users must know they are interacting with AI, and AI-generated content must be identifiable as such. Art.

13 obligations target providers of high-risk systems, while Art.

50 obligations apply to providers and deployers of chatbots, generative AI systems, and deepfake tools.

Violations of transparency obligations under Article

50 can result in fines of up to EUR

15 million or 3% of global annual turnover, whichever is higher. For serious violations involving prohibited AI practices (Art. 5), penalties increase to EUR

35 million or 7% of turnover. In Germany, enforcement will be handled by the market surveillance authority designated under the KI-MIG (AI Implementation Act).

Yes, Article 50(4) provides exceptions: Deepfakes do not need labeling when they are obviously satirical, artistic, or fictional and recognizable as not depicting reality for an average viewer. Content that undergoes editorial review where a natural or legal person assumes responsibility may also be exempt. Special rules apply for law enforcement purposes and certain governmental uses.

Providers of AI systems must ensure synthetic content is identifiable as AI-generated in a machine-readable format. Methods include watermarks, metadata embedding, and cryptographic techniques. The EU AI Office is currently developing a Code of Practice that establishes technical standards for labeling — the final version is expected in June 2026. Labeling must be effective, reliable, interoperable, and state-of-the-art.

The EU AI Act distinguishes clearly between providers (developers) and deployers (users) of AI systems. Providers of generative AI systems must deliver the technical infrastructure for machine-readable labeling. Deployers who use AI systems are responsible for visible disclosure to end users — such as chatbot labeling on their website or deepfake labeling when publishing content. Both sides bear obligations.

The EU AI Act explicitly recognizes trade secret protection (Recital 78). Companies can disclose confidential information to authorities without it becoming public. Instructions for use of high-risk AI under Art.

13 must contain relevant information but need not reveal proprietary algorithms in detail. In practice, this means: transparency about how a system works and its limitations yes, disclosure of source code no.