AI Provider Transparency

Legal Framework

The transparency obligations for AI providers and deployers are primarily governed by Article 50 of the AI Act, which establishes three distinct duties. First, providers must ensure that AI systems designed for direct interaction with natural persons inform those persons they are interacting with an AI system, unless this is obvious from the perspective of a reasonably informed, observant, and circumspect person. Second, providers of AI systems generating synthetic audio, image, video, or text content must mark outputs in a machine-readable format detectable as artificially generated or manipulated. Third, deployers of emotion recognition or biometric categorisation systems must inform exposed individuals of the system's operation.

Article 50's obligations are directly applicable — the regulation does not require national implementing legislation. The Commission is tasked under Article 96(1)(d) to develop practical implementation guidelines specifically addressing these transparency obligations, taking due account of the state of the art and relevant harmonised standards.

Enforcement is significant: non-compliance with Article 50 falls under Article 99(4), which provides for administrative fines of up to EUR 15 million or 3% of worldwide annual turnover, whichever is higher.

The synthetic content marking obligation is qualified by feasibility and technical limitations:

This means providers must balance effectiveness against implementation costs and the generally acknowledged state of the art, as reflected in relevant technical standards.

Key Developments

The Commission's guidelines under Article 96 are still forthcoming, leaving many practical questions unresolved. The EDPB has signalled active interest in the interplay between the AI Act and data protection frameworks, particularly regarding transparency in political advertising. Stakeholders have emphasized the need for clarity on how AI Act transparency obligations interact with GDPR, DSA, and sector-specific rules.

"Stakeholders emphasized the need for visual examples of targeting criteria, specifically to clarify whether TTPA transparency requirements must be applied to each individual criterion or to the campaign"
— EDPB Stakeholder Report §17

This reflects a broader uncertainty: stakeholders agree transparency is important but diverge on what it concretely requires. No court rulings yet interpret Article 50's thresholds, meaning the "reasonably well-informed, observant and circumspect" standard for AI interaction disclosure and the "technically feasible" standard for content marking remain undefined in practice.

Status of the Debate

This topic is actively contested and regulator-defined rather than court-settled. The doctrinal status signals divergent judicial approaches, but no court split is yet on record regarding Article 50 specifically. The boundaries of the transparency obligations — particularly the "obvious from the point of view of a natural person" exception in paragraph 1, the "assistive function" exclusion in paragraph 2, and the scope of "technically feasible" — are open questions awaiting Commission guidelines and eventual enforcement decisions. The EDPB's forthcoming work programme indicates guidance on AI and transparency is in development, which will likely shape the practical thresholds. Resolution will come through a combination of Commission implementing guidelines, harmonised technical standards under Articles 40-41, and initial enforcement actions by national market surveillance authorities.

Practical Guidance

• Interaction disclosure: Design AI systems intended for direct human interaction with built-in disclosure mechanisms, unless the AI nature is genuinely obvious to a reasonably informed user in the specific use context. Document the contextual assessment supporting any non-disclosure decision.
• Synthetic content marking: Implement machine-readable watermarking or metadata tagging for all AI-generated audio, image, video, and text outputs, ensuring interoperability with emerging CEN-CENELEC and ISO/IEC standards. Prioritise robustness against removal or tampering.
• Emotion and biometric systems: Deployers must proactively inform exposed individuals before or at the time of system operation. Establish clear notification protocols and maintain records of compliance.
• Feasibility documentation: For synthetic content marking, maintain technical documentation assessing feasibility, implementation costs, and state-of-the-art limitations, as Article 50(2) explicitly requires these factors to be considered.
• Monitor Commission guidance: Track Article 96 implementing guidelines and harmonised standards development, as these will define the concrete technical and procedural benchmarks for compliance.