ATLAS
← Back to News and Insights

Commission sets non-binding guidance for four AI Act transparency duties

5 min

The guidance requires clear first-contact notices, machine-readable output marks and accessible disclosures for higher-risk interactions.

LinkedIn Post
Commission sets non-binding guidance for four AI Act transparency duties
European Union · Photo: François Genon / Unsplash

The European Commission has published Guidelines on the transparency obligations in Article 50 of the AI Act, Communication C/2026/6005, in the Official Journal on 12 November 2026. Issued under Article 96(1)(d), the Guidelines explain how providers and deployers should apply the four transparency duties covering interactive AI, synthetic content, emotion recognition and biometric categorisation, deepfakes and certain public-interest text.

The Guidelines are expressly “non-binding”. The binding duties come from Article 50, which has applied since 2 August 2026. The Commission says only the Court of Justice of the European Union can give an authoritative interpretation of the AI Act.

The immediate compliance issue is allocation of responsibility. Providers must build transparency into interactive and generative systems before placing them on the Union market or putting them into service. Deployers are responsible for informing people exposed to emotion recognition or biometric categorisation systems, and for labelling deepfakes and certain AI-generated text. Information must be clear, distinguishable and accessible no later than the first interaction or exposure.

The duties can apply cumulatively. A company offering an image-generating chatbot under its own name may therefore have provider duties under Articles 50(1) and 50(2), while the organisation using it to publish a deepfake may also have deployer duties under Article 50(4).

Four duties, two responsible actors

Interactive AI. Providers of systems intended to interact directly with natural persons must inform users that the counterpart is artificial and non-human. The Guidelines describe direct interaction as a genuine, bidirectional exchange, including through chatbots, voice assistants, AI agents, AI companions, social-media bots and coding agents.

A system that merely recommends content, filters spam, translates or transcribes, retrieves information, or provides backend decision support without allowing direct interaction is generally outside Article 50(1). Human review does not by itself avoid the duty where the AI system remains the direct interlocutor.

The disclosure must be specific and perceptible. Terms and conditions, documentation or a URL alone are insufficient. Machine-readable metadata or an invisible watermark does not satisfy the user-facing disclosure requirement, although those tools may support Article 50(2) compliance.

A single prominent notification before the first interaction will usually be enough. The Commission says periodic or context-aware reminders are likely to be needed where users are vulnerable, interactions are sustained or immersive, or there is a heightened risk of deception, including financial advice, insurance, legal assistance, health advice, complaints handling and AI companions. Notices for children should be age-appropriate and accessible.

AI agents receive particular attention. They should disclose both their artificial nature and the person or organisation on whose behalf they act. Where a provider cannot reliably know in advance whether an agent will interact with people, the agent should be designed to disclose itself whenever such interaction is reasonably likely, including at key authorisation, reporting and validation stages.

Synthetic content. Providers of systems generating or manipulating synthetic image, video, audio or text must ensure outputs carry machine-readable markings and are detectable as artificially generated or manipulated. The technical solutions must be “effective, interoperable, robust and reliable”.

The duty does not apply where the system performs an assistive function for standard editing or does not substantially alter the input data or its semantics. There is also an exception for systems authorised by law for criminal-offence detection, prevention, investigation or prosecution.

The Guidelines do not impose Article 50 duties directly on general-purpose AI models. However, the Commission encourages model providers to implement identification and marking measures at model level, particularly where this helps downstream system providers meet their obligations.

Emotion recognition and biometric categorisation. Deployers must inform natural persons exposed to these systems that the systems are operating. The exception covers systems permitted by law to detect, prevent or investigate criminal offences.

Deepfakes and public-interest text. Deployers must disclose that deepfake content has been artificially generated or manipulated. A special disclosure regime applies to deepfakes forming part of artistic, creative, fictional, satirical or analogous works or programmes.

Deployers must also disclose AI-generated or manipulated text published to inform the public about matters of public interest. The obligation does not apply where the text has undergone human review or editorial control and is subject to editorial responsibility. A separate criminal-law exception also applies.

Scope follows control and distribution

The Guidelines treat a provider as the entity developing an AI system, or having it developed, and placing it on the Union market or putting it into service under its own name or trademark. That applies whether the system is free or paid for, and whether the provider is established inside or outside the Union.

A provider outside the Union is covered where the system's output is used in the Union. The Guidelines caution that incidental, unforeseeable or unauthorised downstream use alone should not trigger the obligations for a third-country provider that did not place the system on the market or put it into service in the Union.

For deployers, the relevant question is authority over the decision to use the system and how it is used, not technical control. A company remains the deployer when contractors or freelancers operate the system on its behalf. A third-country deployer is covered where it foresees dissemination or use of the output in the Union, including by posting a deepfake on the globally accessible internet.

By contrast, hosting services, online platforms and broadcasters that merely transmit or disseminate third-party AI content are not deployers under Article 50. The Commission nevertheless strongly encourages them to preserve markings and implement proportionate, technically feasible measures that help people detect artificial content. Their duties or liability under the Digital Services Act and other law remain unaffected.

The personal-use exception is narrow. Individuals using AI for purely personal, non-professional activity are outside the deployer obligations, but the provider of the system must still comply. Open-source AI systems remain subject to Article 50 where the system itself falls within the provision. Open-source components that are not themselves AI systems do not.

In Atlas's view, the structural change is that Article 50 compliance cannot be treated as a disclosure policy added after launch. Providers need interaction notices, provenance controls and accessibility features in the product architecture, while deployers need controls over publication, distribution and downstream interfaces. Compliance also does not make an otherwise prohibited or high-risk use lawful, and does not replace AI literacy, high-risk or other AI Act duties.

The practical date for these controls is already 2 August 2026. The Commission's 12 November Guidelines clarify how to apply them, but do not create a new commencement period.

Official source: eur-lex.europa.eu
AIAI FrameworksAI Governance

Permanent link to this Atlas analysis

Continue with coverage connected by market, topic and operating context.

Continue with Atlas

Move from this development into the relevant research, comparison and workflow.