The EU AI Act imposes transparency obligations on providers and deployers of certain AI systems placed on the market or used in the EEA. From 2 August 2026, businesses providing or using these AI systems will have to ensure that people know when they are interacting with AI or are exposed to certain AI-generated or manipulated content. The European Commission’s new guidelines explain which organisations are responsible for each obligation and what they need to do in practice. They also complement the voluntary Code of Practice which focuses on transparency of AI-generated content, published in June.
The guidelines are divided into sections, providing an explanation of the distinction between providers and deployers and considering the requirements applicable to interactive AI systems, AI-generated content, emotion recognition, biometric categorisation and deepfakes. They also include a separate section on horizontal requirements governing how the information required under Article 50 must be provided.
Overview
The transparency obligations apply differently to providers and deployers of AI systems, so the first section of the guidelines helps to distinguish between these roles. Providers are persons or organisations that develop AI systems, or have them developed, and put them on the market or into service in the EEA under their own name or trade mark. Deployers are persons or organisations that use AI, unless this is for a personal non-professional activity. The guidelines explain how an organisation may fulfil more than one role concurrently, for example if it develops an in-house generative AI system and uses it to generate deepfakes.
This section also explains the key exceptions to the transparency obligations, including when use of an AI system is considered personal and non-professional.
Finally, it explains how the rules on general-purpose AI (GPAI) models and systems interact with Article 50 of the AI Act. Article 50 does not explicitly apply to GPAI models, but it does apply to certain GPAI systems, eg chatbots that interact directly with individuals or AI systems that generate or manipulate synthetic content. Where an interactive or generative AI system that falls within the scope of Article 50(1) or (2) of the AI Act is built upon a GPAI model provided by the same provider, the transparency measures set out in the guidelines could also be implemented at the model level. Furthermore, while GPAI models fall outside the scope of Article 50, providers of other GPAI models and generative AI models are also encouraged to implement appropriate transparency measures at the model level for the identification of interactive AI systems and the marking of AI-generated and manipulated content.
Article 50(1): Transparency for interactive AI systems
This paragraph obliges providers of AI systems directly interacting with individuals to design and develop their systems so that the individuals are informed that they are interacting with an AI system, at the latest at the time of their first interaction. The guidelines explain which systems fall within this paragraph’s scope, providing examples. For example, AI companions and robotic companion pets are in scope, but AI-enabled recommender systems are not, because they do not involve the required exchange of information with a conversational or responsive character.
Article 50(1) provides for two exceptions to this obligation: AI systems authorised for law enforcement purposes, and where it is obvious to the average consumer that they are interacting with an AI system. The guidelines provide examples of where this is obvious, for example an internal AI assistant for properly-trained, AI-literate staff, and where it is not obvious, for example AI chatbots on online platforms where users receive AI outputs they may perceive as human-generated.
Article 50(2): Marking and detection of AI-generated or manipulated content
This paragraph obliges providers of AI systems generating synthetic content to implement technical solutions to ensure that the system’s outputs are marked in a machine-readable format and detectable as artificially-generated or manipulated. The guidelines explain which systems this obligation applies to, including virtual reality and digital twins, and the exceptions, including industrial and business-to-business AI applications, and systems used for standard editing, eg spelling and grammar correction. They then break down the requirements into the marking obligation, the detection obligation, and the requirement that the technical solutions implemented must be effective, interoperable, robust and reliable.
Article 50(3): Emotion recognition systems and biometric categorisation systems
This paragraph obliges deployers of emotion recognition systems and biometric categorisation systems to inform individuals who are exposed to those systems of their operation, at the latest at the time of their first interaction with or exposure to the AI system. The guidelines flag that, because these systems are classified as high-risk, unless they are prohibited, the (separate) Commission guidelines on the classification of high-risk AI systems also apply.
Article 50(4): Labelling of deepfakes and AI-generated or manipulated text
This paragraph imposes obligations on deployers of generative AI systems requiring them to disclose that certain content has been AI-generated or manipulated. The guidelines break these down into the separate obligations applicable to the relevant categories of content:
Deepfakes
The guidelines explain what the term covers and provide examples of content that does and does not constitute a deepfake. For example, an advertisement showing an AI-generated image of a product that could mislead the audience as to the real product’s appearance, characteristics or use would be considered a deepfake. In contrast, an ad showing a real product against an AI-generated background would not be considered a deepfake, provided that it is not likely to mislead the audience about the product or its characteristics and use.
This section also explains the disclosure obligation and the limited obligations that apply to artistic, creative, satirical, fictional or analogous deepfake content.
AI-generated or manipulated text published with the purpose of informing the public on matters of public interest
The guidelines explain the meaning of this term, including matters of public interest, and the exception for text under human review or editorial control and editorial responsibility.
Article 50(5): Horizontal requirements applicable to information provided under Article 50(1)-(4)
This paragraph provides that the information required under Article 50(1)-50(4) must be provided in a clear and distinguishable manner, at the latest at the time of the first interaction or exposure, and in compliance with applicable accessibility requirements.
Enforcement
This section seeks to clarify several points relating to enforcement, including:
- Codes of practice: How providers and deployers of AI systems within the scope of Article 50(2) and 50(4) of the AI Act can demonstrate compliance with their transparency obligations by adhering to a code of practice, and alternative ways of demonstrating compliance.
- Penalties for breach: fines of up to €15 million or 3% of worldwide annual turnover for organisations other than EU institutions and SMEs, which are subject to lower fines.
- Entry into application: the transparency obligations become applicable on 2 August 2026, but the AI Omnibus has introduced a grace period until 2 December 2026 for systems placed on the market or put into service before 2 August 2026 to implement the technical marking and detection measures required for AI-generated content.
Next steps
Businesses should identify the AI systems they provide or use, determine their role in relation to each AI system, and assess which transparency obligations apply. They should also review user notices, technical marking solutions, content-labelling processes and internal human-review arrangements, and document the approach taken to compliance before the rules apply on 2 August 2026.
