European UnionEuropean Commission Guidelines on the implementation of the transparency obligations for certain AI systems under Article 50 of the AI Act
Guidance, not a law: the European Commission's reading of AI Act, Article 50 (transparency obligations for AI systems and synthetic content). It binds nobody by itself; the law it reads does.
Guidance on an AI transparency rule, addressed to public and private bodies.
- Private right of action
- No
- Instrument type
- guidance published by a regulator
- Obligation class
- Content labelling, Disclosure
As of .
What the regulator expects
- If you are a provider of an AI agent that is capable of interacting with natural persons, design and develop it so that it discloses both its artificial nature and the person on whose behalf it acts (paragraph 31). Where you cannot reliably determine in advance whether the agent will interact directly with a natural person, the guidelines say the agent should be designed at the architecture level, and instructed, to disclose itself as such in every situation where it is reasonably likely to interact with a natural person (paragraph 31).NIST AI RMFMEASURE 2.8
- If your product or service blends AI-generated responses with human-curated content, disclose the specific outputs in the interaction that are AI-generated, unless those outputs have been properly reviewed and sent by humans as the main interlocutors with the natural persons; the guidelines say the mere possibility for humans to intervene or review the outputs should not be used to circumvent the disclosure requirement (paragraph 30).NIST AI RMFMEASURE 2.8
- Inform the people who interact with your AI system explicitly of the artificial, non-human nature of the counterpart, at the latest at the time of the first interaction (paragraphs 33 and 35). The guidelines say that disclosures contained only in terms and conditions, URLs or documentation, machine-readable markings not perceivable at the point of interaction, generic references to an “assistant”, and generalised statements such as “Services on this website use AI” are insufficient when used alone (paragraph 38). They also say that periodic reminders and context-aware disclosures are likely to be necessary in riskier contexts, such as interaction with vulnerable persons, AI companions, financial, insurance, legal or health advice, and AI agents, and that disclosure must be ensured whenever the system is asked about its nature or the person is likely to be misled about the AI origin of the interaction (paragraph 40).NIST AI RMFMEASURE 2.8
- To rely on the exception for an interaction whose artificial origin is obvious, assess and demonstrate that the artificial nature is obvious to a reasonably well-informed, observant and circumspect natural person in the audience, taking into account the circumstances and the context of use (paragraph 42). The guidelines say the exception should be interpreted restrictively and limited to cases where almost no doubt is left about the nature of the interaction, and that it cannot be relied on where the audience may include vulnerable persons for whom the interaction is less obvious (paragraph 45). The law-enforcement exception does not cover an AI system that is available to the public and offers a functionality to report criminal offences (paragraph 49).NIST AI RMFMEASURE 2.8
- If you are a provider of a generative AI system, implement both a machine-readable marking of the outputs and the means to detect them as AI-generated or manipulated: the guidelines say that fulfilling only one element, such as marking without available detection, will not suffice (paragraphs 69 and 70). Where no publicly available industry standard detection solution exists, the guidelines say you may rely on your own, a third party or a shared detection solution so long as it ensures interoperability with the detection solutions used by other providers, and that this possibility should be limited in time until harmonised standards emerge (paragraph 76).NIST AI RMFMEASURE 2.8
- Ensure that the combination of your marking and detection solutions holistically meets the requirements of effectiveness, reliability, robustness and interoperability, insofar as this is technically feasible (paragraph 80). The guidelines say technical feasibility is an objective notion that does not depend on the resources and capabilities of an individual provider, and that you must continuously adapt your solutions as the state of the art evolves (paragraphs 81 and 83).NIST AI RMFMEASURE 2.8
- If your system generates or manipulates audio, image, video or text, mark outputs perceptible to natural persons, including AI-agent actions whose output is AI-generated content perceptible to them (paragraph 63). The guidelines place outside Article 50(2) outputs that are a short sequence of numbers, symbols or letters, source code, outputs exclusively communicated from machine to machine without exposure to humans, and outputs used only in closed loop industrial and product development workflows, where only the final output from workflow stages has to be marked and detectable (paragraph 68).NIST AI RMFMEASURE 2.8
- Mark content whose meaning is changed by the AI system: the guidelines treat AI-generated summaries of text, paraphrasing that changes style, structure and meaning, face replacement or substantial facial modification, and the synthesis of realistic speech in a specific person's voice as semantic changes that require marking, and treat grammar correction, spellchecking, AI-generated translations of text, formatting and minor color or cropping adjustments as standard editing or minor alterations that benefit from the exception (paragraph 92).NIST AI RMFMEASURE 2.8
- If you are a deployer of a biometric categorisation system or an emotion recognition system, inform every natural person exposed to its operation, including children, that the system is operating, whether they are exposed in real time or the system is operated ex post (paragraphs 100 and 106). The guidelines say Article 50(3) applies to any biometric categorisation system, such as age or gender classification based on biometric data, unless it is prohibited under Article 5(1)(g), whether or not it is high-risk, and that the notice need not give the reasons for the operation of the system or any other purpose of processing, which Union data protection law may require separately (paragraphs 104 and 105).NIST AI RMFMEASURE 2.8
- If you are a deployer of an AI system that generates or manipulates deep fakes, label them with a label understandable and perceivable by natural persons, such as a visible or audible label, without any specific technical tool or dedicated action; the guidelines say you cannot rely on the machine-readable marking embedded by the provider (paragraph 117). The guidelines say whether content would falsely appear authentic or truthful is assessed objectively and does not require an intention to deceive, taking into account the intended and reasonably foreseeable audience, including children and the elderly (paragraphs 114 and 115).NIST AI RMFMEASURE 2.8
- If you rely on the limited disclosure for a deep fake that forms part of an evidently artistic, creative, satirical, fictional or analogous work or program, still disclose its artificial origin in an appropriate manner that does not hamper the display or enjoyment of the work (paragraph 123). The guidelines say the categories are interpreted strictly, that content whose nature is unclear or exclusively informative or commercial is excluded, that the informative character prevails where a deep fake combines several characters, and that you must still safeguard the rights and freedoms of third parties (paragraphs 122 and 124).NIST AI RMFMEASURE 2.8
- If you are a deployer of an AI system that generates or manipulates text published to inform the public on matters of public interest, label the text as artificially generated or manipulated unless it has undergone human review or editorial control and a natural or legal person holds editorial responsibility for the publication (paragraphs 132 and 133). The guidelines say human review includes fact-checking as a minimum, that superficial, solely formal or procedural checks such as spell-checking do not qualify, that substantive AI intervention after editorial sign-off voids the exception, and that the identity and contact details of the person or function with editorial responsibility should be made publicly available in an easily findable location (paragraphs 134, 135, 136 and 138).NIST AI RMFMEASURE 2.8
- Provide the information required by Article 50(1) to (4) in a clear and distinguishable manner, so that it cannot be easily overlooked or missed under normal exposure or interaction conditions, for example by placing it only in a manual, behind menu options or in terms of use (paragraph 142). The guidelines say the first-exposure timing applies to every natural person who is first exposed, that for AI-generated content the information obligation applies to each output, and that disclosure only at the beginning of content should be complemented with later disclosure where it is reasonably foreseeable that persons will not perceive the content from its beginning (paragraph 143).NIST AI RMFMEASURE 2.8
- If you do not sign a code of practice that is assessed as adequate under Article 50(7), be ready to demonstrate through other adequate means how you comply with Article 50(2), (4) and (5), and to explain how your measures ensure compliance; the guidelines say you should, for instance, carry out a gap analysis against the measures of an adequate code, and that you may be subject to a larger number of requests for information and for access to assess the effectiveness, interoperability, robustness and reliability of your technical solutions (paragraph 148).NIST AI RMFGOVERN 1.1
In the frameworks
- NIST AI RMF
- GOVERN 1.1Legal and regulatory requirements involving AI are understood, managed, and documented.
- MEASURE 2.8Risks associated with transparency and accountability – as identified in the MAP function – are examined and documented.
- NIST AI 600-1
- GAI-RISK-07Human-AI Configuration
- GAI-RISK-08Information Integrity
- OWASP Agentic Top 10
- ASI09Human-Agent Trust Exploitation
- MIT mitigations Draft
- NIST Privacy Framework
- CM.AW-P1Mechanisms (e.g., notices, internal or public reports) for communicating data processing purposes, practices, associated privacy risks, and options for enabling individuals’ data processing preferences and requests are established and in place.
- FINOS AIGF
- AIR-DET-021Agent Decision Audit and Explainability
- ETSI EN 304 223
Our reading of where this law's duties sit in each framework, and in the NIST AI RMF, line by line beside each duty above. It never says that running a control meets the law. How we read a framework against the law
Identifiers refer to ETSI EN 304 223 V2.1.1 (2025-12), © ETSI 2025, all rights reserved. ETSI owns the standard; no ETSI text (titles or sentences) is reproduced here. The standard is free to download from ETSI.
Who enforces it
Enforcement body
National market surveillance authorities designated by the Member States, the AI Office for AI systems within its competence, and the European Data Protection Supervisor for the Union institutions, bodies and agencies; the guidelines carry no enforcement action of their own.
Settledness
The guidelines are the Commission's own reading of Article 50 and are issued as practical guidance, so their own text is the guidance on the duty.
- As of
- Open questions
- Do the guidelines apply as guidance to providers and deployers before the Commission formally adopts them in all language versions, given that the Communication approving their content says it is only from formal adoption that they will be applicable?
What it makes you log
Logging duty
The guidelines say that providers are not required to record or keep a full provenance chain containing information on content origin and modifications, and they state no duty to keep a log of disclosures or markings.
- Kind
- None
- As of
What this law does
The Communication to the Commission C(2026) 5054 final, dated , approves the content of the draft guidelines and says that the guidelines will be formally adopted by the Commission at a later date, when all language versions are available. The same Communication says that it is only from the moment of formal adoption that the guidelines will be applicable.
Paragraph 5 of the guidelines says that they are non-binding and that any authoritative interpretation of the AI Act may ultimately only be given by the Court of Justice of the European Union. Paragraph 3 says that the guidelines are issued pursuant to Article 96(1)(d) of the AI Act as practical guidance to assist competent authorities, providers and deployers in ensuring compliance with Article 50.
Paragraph 152 says that providers and deployers that do not comply with the applicable transparency obligations laid down in Article 50 may be fined up to EUR 15,000,000 or, for an undertaking, up to 3% of total worldwide annual turnover for the preceding financial year, whichever is higher. Paragraph 153 says that Article 50 applies from .
Paragraph 153 also says that a system that is partly interactive and partly generative may use the transition to only for the Article 50(2) marking obligation, and that the disclosure obligation for systems directly interacting with natural persons must be ensured as of . Paragraph 146 says that providers and deployers may demonstrate compliance with Article 50(2) and (4) by adhering to a code of practice that is assessed as adequate pursuant to Article 50(7).
Paragraph 146 adds that such a code does not replace the AI Act or the guidelines. Paragraph 147 says that, for signatories of a code assessed as adequate, the Commission and the competent market surveillance authorities will focus their supervisory activities on whether those signatories have adhered to the code and implemented its measures.
The Commission's page on its opinion says that on the Commission concluded that the Code of Practice on Transparency of AI-generated Content adequately covers the obligations in Article 50(2), (4) and (5). The same page says that the AI Board adopted its adequacy assessment of the code the following day. The Commission's questions and answers page on Article 50 describes itself as answering questions on concepts and obligations clarified in the guidelines.
Paragraph 155 says that the guidelines are a first interpretation and that, on a review, the Commission may decide to withdraw or amend them.