On 20 July 2026, less than a fortnight before the transparency obligations in Article 50 of the EU AI Act become applicable, the European Commission published its Guidelines on the implementation of the transparency obligations for certain AI systems under Article 50 (C(2026) 5054). The Guidelines are the Commission’s first comprehensive interpretation of the provision that will, from 2 August 2026:

• determine how chatbots and AI agents identify themselves,
• how AI-generated content is marked and detected, and
• how deepfakes and certain AI-generated text publications are labelled.

Together with the Code of Practice on Transparency of AI-Generated Content published on 10 June 2026, they complete the guidance framework for the first AI Act obligations most organisations will have to meet.

The Guidelines are issued under Article 96(1)(d) AI Act to assist providers, deployers and competent authorities in applying Article 50 consistently. They are non-binding, and the authoritative interpretation of the AI Act ultimately rests with the Court of Justice. Yet, they will frame supervisory expectations from the outset, and enforcement begins at national and EU level on 2 August 2026 with fines of up to €15 million or 3% of total worldwide annual turnover, whichever is higher.

Deadline Looming

Article 50’s transparency obligations apply from 2 August, 2026.

However, AI systems, including general-purpose AI systems, that generate synthetic audio, images, video, or text (“generative AI systems”) placed on the market before 2 August 2026 have until 2 December 2026 to comply with the machine-readable marking obligation in Article 50(2).

Systems that are partly interactive and partly generative also benefit from the transitional period, but only with regard to the requirement to be marked as AI-generated.

The Guidelines also confirm that content generated before 2 August 2026 need not be marked or labelled retroactively, although AI-generated text published on or after that date must be labelled even if it was generated earlier.

Interactive AI systems and AI agents

Article 50(1) requires providers to design AI systems intended to interact directly with natural persons so that those persons are informed they are dealing with AI. The Guidelines break the scope into four elements (an AI system intended to interact directly with natural persons) and give the provision a decisively agentic reading. AI agents must disclose both their artificial nature and the person on whose behalf they are acting. Where a provider cannot reliably predict whether an agent will encounter natural persons once deployed, the agent must be designed at the architecture level to disclose itself in every situation where interaction is reasonably likely. The obviousness exception is to be interpreted restrictively by reference to the average consumer standard, and the Guidelines state that general public awareness that chatbots exist does not mean people recognise them in individual interactions. Disclosures buried in terms and conditions, generic references to an “assistant”, and machine-readable marks that users cannot perceive are all identified as insufficient on their own.

Marking and detection of synthetic content

Article 50(2) contains two interlinked obligations, and the Guidelines give the second real prominence. Providers of generative AI systems must mark outputs in a machine-readable format and must ensure that means of detection are available to those exposed to the content; marking without available detection will not achieve compliance. Detection should rely on publicly available industry-standard solutions, ideally locally executable, and providers may use their own or third-party detection solutions only until harmonised, provider-agnostic standards emerge. Marking may be implemented at the level of the AI system, the underlying model or by a third party. Providers may rely on the solution implemented by an upstream model provider, although responsibility for demonstrating compliance remains with the system provider. This confirms the contractual flow-down strategy: downstream providers need warranties and evidence that upstream marking is present, persistent and compliant.

The Guidelines also draw practical boundaries around the obligation. Source code, short outputs such as captions and interface labels, purely machine-to-machine communications and intermediate outputs in closed production workflows fall outside scope, with only final outputs requiring marking. A narrow exemption covers strictly technical industrial and business-to-business outputs confined within the organisations concerned. Ephemeral real-time content in video games and virtual reality may be exempt where marking is technically infeasible and users are informed in-experience. On the standard editing exception; grammar correction, translation and technical enhancement fall outside the obligation, while AI-generated summaries and paraphrasing that changes meaning, style or structure must be marked.

Deepfakes and public-interest text

For Article 50(4), the Guidelines dissect the deepfake definition in Article 3(60) into four cumulative criteria:

• appreciable resemblance to the subject
• a subject that exists or could plausibly exist in reality
• the listed categories of persons, objects, places, entities or events, and
• the capacity to appear falsely authentic or truthful to the reasonably foreseeable audience.

Evidently, fantastical content falls outside the definition. Critically, deployers cannot discharge the labelling duty by pointing to the provider’s machine-readable marks: labels must be perceivable by natural persons without technical tools or dedicated actions. The attenuated regime for evidently artistic, creative, satirical or fictional works is to be construed strictly, and where content combines informative and creative characters, the informative character prevails and, full labelling applies.

On AI-generated text published to inform the public on matters of public interest, the Guidelines set a high bar for the editorial carve-out.

Human review means substantive examination of the content, including fact-checking;

• Superficial, formal or automated checks will not qualify;
• Any substantive AI intervention after editorial sign-off voids the exception;
• and the natural or legal person holding editorial responsibility must be publicly identifiable.

Organisations publishing AI-assisted content will need documented editorial workflows with named responsibility.

Emotion recognition and biometric categorisation

Article 50(3) requires deployers of emotion recognition and biometric categorisation systems to inform the natural persons exposed to them of the system’s operation, in addition to any disclosures required under data protection law. The Guidelines confirm that the notice is required whether the system operates in real time or ex post, and must be given at the latest at the time of first exposure.

Demonstrating compliance

The Guidelines formalise the compliance value of the Code of Practice. For signatories of a code assessed as adequate, supervision will focus on whether the code’s measures have been implemented, and commitments implemented under the code may be taken into account as a mitigating factor when fines are set. Providers and deployers choosing another route are expected to demonstrate compliance through equivalently adequate means, including a gap analysis against the code, and should expect more detailed information requests from market surveillance authorities. In Ireland, supervision will sit with the sectoral regulators designated under the European Union (Artificial Intelligence) (Designation) Regulations 2025, coordinated by the AI Office of Ireland under the Regulation of Artificial Intelligence Bill 2026.

With less than two weeks to the application date, the workstream is unchanged and now urgent:

• an inventory of every interactive, generative, emotion recognition and biometric categorisation system reaching EU users;
• classification of the organisation’s role as provider or deployer for each;
• disclosure design against the clarity, timing and accessibility requirements of Article 50(5);
• a marking and detection plan for generative AI systems;
• contractual measures covering upstream marking warranties,
• detection access, and restrictions on removing marks or labels;
• a documented record of the decisions taken.

The Commission has signalled that it will review the Guidelines as enforcement experience accumulates. The high-risk deferrals bought organisations time, and the transparency obligations have run out of it.