Although the Guidelines are non-binding, they are intended to shape how national market surveillance authorities enforce the rules across the EU. 

The Guidelines are designed to be read alongside the Code of Practice on Transparency of AI-Generated Content (the “Code”), which the Commission and the AI Board have concluded is an adequate voluntary tool to demonstrate compliance with the transparency obligations. Signing up to the Code is not compulsory, but it is currently the only EU-wide framework the Commission has endorsed for demonstrating compliance with the EU AI Act’s marking and labelling obligations. Non-signatories may expect closer scrutiny.

Overview of the Guidelines

The Guidelines address each of the four Article 50 transparency requirements in turn, from Article 50(1) to Article 50(4), before setting out the overarching rules in Article 50(5). Those cross-cutting rules require information to be provided in a “clear and distinguishable manner”, at the point of first interaction or exposure (at the latest), and in an accessible format. The obligations apply cumulatively: a single AI system may engage more than one limb, and compliance obligations may fall on both the provider and the deployer. We consider each obligation in detail below.

Provider Obligation: Interactive AI (Article 50(1))

Article 50(1) applies to providers of AI systems intended to interact directly with individuals. Examples include chatbots, voice assistants, avatars, humanoid robots, companion devices and AI agents. The Guidelines set four cumulative criteria for a system to fall within Article 50(1):

1. it must fulfil the definition of an AI system (explicitly excluding non-AI automation such as rule-based quick answers);

2. it must be intended to interact with natural persons (via auditory, visual or physical form);

3. the interaction between the system and the person must be direct (specified as real-time or near real-time interaction); and

4. the interaction must be with a natural person (not a backend machine-to-machine call). 

The Guidelines further state that, to comply, providers should embed the disclosure into the system’s design so that it appears at the very first interaction. Examples include a plain-language notice at the start of a session, a persistent badge near the input field, or a spoken statement in a voice interaction. Generic references to an “assistant”, purely machine-readable marks, or disclosures buried in terms and conditions will not suffice.

Where an AI agent acts autonomously, it must also disclose on whose behalf it is acting, and reaffirm this at key steps such as authorisation and reporting. Extra care is expected where the audience may include children, elderly users or people with lower digital literacy.

Provider Obligation: Marking Generated Content (Article 50(2))

Article 50(2) requires providers of generative AI to mark synthetic image, audio, video and text outputs in an effective and reliable machine-readable format. The Guidelines note that providers should meet the generally accepted “state of the art” standard, but must also take proportionality and cost into account when applying these requirements. 

The Guidelines set out a number of carve-outs from Article 50(2). These include short outputs such as alt-text, captions and interface labels; source code (broadly defined); AI-generated translations; and standard editing that does not change substance or meaning. Outputs used in closed-loop environments are also excluded, provided that the output is not shared outside the organisation and is protected by safeguards such as access controls. Public and consumer-facing AI is not eligible for these carve-outs.

The Guidelines also note that a downstream provider may rely on marking implemented at model level by an upstream general-purpose AI provider, but must still demonstrate that the solution actually works. Providers should therefore document their technical choices and revisit them as marking technology evolves.

Deployer Obligation: Emotion Recognition and Biometric Categorisation (Article 50(3))

Moving to the deployer requirements, Article 50(3) applies to deployers of emotion recognition or biometric categorisation systems. Examples of these systems include:

• a wearable that infers mood;

• a game capturing a player's expressions; or

• a system that groups visitors by age.

Deployers of such systems must inform individuals of their operation at the latest at the point of first exposure.

The Guidelines note that deployers should provide a visible or audible notice calibrated to the environment, such as an onboarding pop-up before a game launches, signage where facial images are captured, or standardised icons where these will be recognised. This notice sits alongside existing UK and EU data protection information duties, and the two can often be combined.

Deployers should note that compliance with Article 50(3) does not, on its own, make the underlying use lawful. Deployers must still comply with any applicable prohibitions under Article 5 of the EU AI Act, such as those relating to emotion recognition in the workplace or in education.

Deployer Obligation: Deep Fakes and Public Interest Text (Article 50(4))

Article 50(4) places two labelling obligations on deployers of generative AI. The first covers deep fakes, and the second covers AI-generated text published to inform the public on matters of public interest. 

The Guidelines provide an overarching definition of what constitutes a deep fake, being AI-generated or manipulated images, audio or video content that resembles an existing person, object, place, entity or event and would falsely appear to be authentic or truthful. The Guidelines then detail a holistic, case-by-case assessment of whether content is a deep fake, requiring assessment of:

• the level of resemblance of the content;

• whether the content resembles existing subjects;

• whether the content manipulates a person, object, place, entity or event; and

• the expectations of the foreseeable audience. 

The Guidelines provide a list of specific examples of what does and does not constitute a deep fake, with a notably lower threshold where the audience includes children or vulnerable groups. 

In every deep fake case, the deployer must apply a clear and distinguishable label at the point of first exposure. The label must be visible or audible and perceivable without special tools. A lighter regime applies to deep fakes forming part of evidently artistic, creative, satirical, fictional or analogous works. However, the Commission construes those categories strictly, and where content mixes informative and creative elements, the informative character prevails. 

In relation to public interest text, the Guidelines describe this broadly as spanning politics, public administration, justice, public health, consumer safety and cultural, financial or scientific developments open to public debate. The editorial control exception for public interest text requires substantive human review, not a cursory check, and any material AI intervention after sign-off will unwind it. Contracts with creative agencies, journalists and downstream distributors should be reviewed so that the labelling obligation follows decision-making.

Next steps

With Article 50 now in force from 2 August 2026, organisations should not assume that their existing AI governance frameworks adequately address the new transparency requirements. The late publication of the Guidelines means that many businesses will need to revisit prior compliance assessments and confirm that their disclosures, labelling measures and marking solutions align with the Commission's latest guidance.

The Guidelines, which the Commission will keep under review, provide a useful starting point for organisations addressing practical questions about the operation of Article 50. 

Given that non-compliance with Article 50 can attract fines of up to €15 million or 3% of global annual turnover, organisations should review the Guidelines carefully to identify:

1. whether any of their AI systems fall within the scope of one or more of the Article 50 transparency obligations; 

2. whether they bear responsibility as a provider, deployer, or both in respect of these systems; and

3. how and where the required disclosures or markings will actually be implemented in practice. 

It should be noted that the transparency obligations are not retrospective. Content generated before 2 August 2026 will not need to be relabelled, but text generated before that date and published on or after it must be labelled unless the editorial control exception applies. Nevertheless, voluntary retrospective labelling of legacy content may represent a best practice approach where transparency concerns are likely to arise. 

Providers of generative AI systems already on the EU market before 2 August 2026 have a short transitional window, until 2 December 2026, before the Article 50(2) machine-readable marking obligations take effect. We discuss this transitional period, along with the other key compliance deadlines reshaped by the Digital Omnibus on AI, in our briefing, "The Digital Omnibus on AI: Ten Key Takeaways". The briefing provides a practical overview of the changes and the steps that organisations should be taking now to prepare for the upcoming implementation dates.

With thanks to Sam Pollock for his contributions to this article.