EU AI Act Article 50: Transparency Requirements

Article 50 addresses transparency for defined AI interactions and AI-generated or manipulated content. The applicable duties differ by role and context, but the shared operational question is simple: can your team show users what they need to know and prove that the control worked in the shipped experience? The Commission states that the relevant Article 50 obligations apply from 2 August 2026.

Where product teams should look first

• AI interaction disclosure — identify journeys where a person interacts directly with an AI system and assess whether the interaction is obvious from the circumstances and context.
• Generated or manipulated content — map text, image, audio, and video generation flows; determine which technical marking, detection, or disclosure duties apply to your role.
• Deepfakes and public-interest content — involve legal and communications owners early where generated or manipulated content could be mistaken for authentic material.
• Supplier dependencies — record what model providers, media-generation tools, and distribution platforms supply so the final disclosure is not based on an assumption.

Evidence worth retaining

Keep the user-flow decision, the disclosure copy and placement, release evidence, model or media-provider version, test results, exceptions considered, and named owner together. When a prompt, model, channel, or audience changes, review the transparency assessment alongside the release change.

How Agent Mai fits

Agent Mai helps product, legal, and security teams keep source evidence and review notes connected to their compliance narrative. It supports structured internal review; it does not determine whether a legal exception or disclosure rule applies to a specific deployment.

Map Article 50 by role and output

Article 50 contains several distinct duties. Providers of systems intended to interact directly with people address interaction disclosure. Providers of systems generating synthetic audio, image, video, or text address machine-readable marking and detectability under the applicable conditions. Deployers address disclosure for deepfakes and certain AI-generated or manipulated text published to inform the public on matters of public interest. The responsible party and control are therefore not the same for every product.

Article 50 implementation checklist

• Inventory direct AI interactions and generated or manipulated text, image, audio, and video flows across product, marketing, support, and internal publication tools.
• Classify the organisation's role for each flow and identify which Article 50 paragraph, exception, or obviousness analysis is relevant.
• Define when the user receives the information, ensuring it is clear, distinguishable, accessible, and presented no later than the first interaction or exposure where required.
• For generated media, document the technical marking or detection method, interoperability assumptions, robustness testing, and any supplier dependency.
• For deepfakes and public-interest text, define editorial ownership, labelling language, review, placement, and evidence retention.
• Add reassessment triggers for model, media pipeline, interface, audience, channel, language, supplier, and legal-guidance changes.

Design disclosures people can actually understand

A legally relevant disclosure should not be buried in terms or shown only after the interaction. Test whether the label is visible at the right moment, understandable in the user's language, available to assistive technology, and preserved across embedded or redistributed content. For conversational systems, product teams should define behaviour when a transcript, screenshot, or output leaves the original interface.

Technical marking and detectability

For relevant synthetic content, providers should assess machine-readable marking at the output layer and whether the method remains effective through common transformations such as resizing, compression, transcoding, editing, screenshots, or metadata removal. Document coverage by content type, the technical standard or method, false-positive and false-negative considerations, interoperability, supplier responsibilities, and test results. A visible label and a machine-readable marker solve different problems and may both be relevant.

Exceptions still need evidence

Article 50 includes context-specific exceptions and qualifications, including authorised law-enforcement uses and certain editorial or artistic contexts. An obviousness conclusion for an AI interaction should be tested from the perspective described in the Act, not assumed because the product team knows how the feature works. Record the exception, facts, reviewer, decision date, and change triggers so the rationale can be revisited when the interface or audience changes.

Code of Practice and non-signatories

The Commission describes the transparency Code of Practice as a voluntary route to demonstrate compliance for relevant marking and labelling duties. Non-signatories remain responsible for compliance and should be prepared to explain how alternative measures provide equivalent support. That makes a documented gap analysis against the Code useful even where an organisation does not sign it.