The European Commission has published Draft Guidelines in relation to the transparency obligations under Article 50 of the EU AI Act (the Act).

The Guidelines are based on stakeholder feedback from a Commission‑led public consultation and input from Member States through the European AI Board, which plays a key role in the governance framework set out by the Act. The Guidelines are non-binding, and the CJEU remains the ultimate interpreter of the Act. However, national authorities and the European AI Office, which ensures coherent application of the AI Act across Member States, will likely adhere to them closely, with any divergence needing to be justified.

Article 50 will be in force from 2 August 2026. The consultation on the Guidelines closed on 3 June 2026. From 2 August 2026, all AI systems within scope of Article 50 that are placed on the market or put into service in the European Union must comply with the transparency obligations, irrespective of when they were first placed on the market or put into service.

Non-compliance with Article 50 may lead to fines up to €15,000,000 and up to 3% of total worldwide annual turnover for the preceding financial year for offending undertakings, whichever is higher. EU institutions, bodies and agencies that violate the transparency obligations could be fined up to €750,000.

Applicable across all of Article 50

The Guidelines include commentary to help navigate and interpret the Article provisions that were initially unclear when the AI Act was published. The following is applicable across the main subsections in Article 50.

• The Guidelines clarify that actors whose role is limited to the dissemination or transmission of AI-generated or manipulated content created by third parties do not qualify as deployers where they do not exercise authority over the AI system. Nevertheless, such actors are encouraged to preserve any marking and labelling applied in accordance with the AI Act.
• Article 2(10) AI Act provides ‘a purely personal, non-professional activity’ exception should be interpreted narrowly. For example, a deepfake depicting a local politician and shared on social media to criticise their decision-making may, at first glance, appear to fall within an exception; however, this would not be the case, as it could influence public discourse on matters of public interest.
• AI systems released under free and open-source licensing remain fully subject to Article 50.
• Article 50 applies to certain AI systems, including those built on GPAI models, such as chatbots or systems that generate or manipulate synthetic content. Article 53 attaches at the model layer. The Guidelines confirm that a generative AI system built on a GPAI model must mark its outputs. GPAI model providers, including those whose models are subsequently integrated into downstream AI systems by third parties, are encouraged to implement content marking and transparency measures directly at the model level, even where those providers are not formally subject to Article 50 obligations.This is intended to facilitate compliance by downstream system providers. In addition, for GPAI models presenting systemic risk, model-level transparency measures may form part of broader mitigation obligations under Article 55 of the AI Act.

Disclosure of AI Interaction and the Narrow “Obviousness” Exception

Article 50(1) AI systems directly interacting with natural persons must be developed in such a way that natural persons are aware they are interacting with an AI system.  Examples provided in the Guidelines include chatbot agents, bots on social networks, and AI-enabled voice assistants. This obligation does not apply where the AI interaction is obvious to a reasonably informed natural user. The Guidelines interpret the “obvious interaction” exception narrowly by requiring providers to assess whether a typical user, considering the audience and context, would clearly recognise the AI nature of the interaction.

The Guidelines explain that assessing the “average” user involves first identifying the intended (and reasonably foreseeable) audience, and then determining how informed, observant and careful a typical member of that audience would be in context. This assessment should consider the presence of vulnerable users, the audience’s level of digital and AI literacy, and relevant social, cultural and linguistic factors.

It is interesting that the Guidelines address agentic AI when the AI Act itself failed to reference it. The Guidelines fill this gap by suggesting that where a provider cannot reliably assess whether an AI agent will interact with a natural person, the agent should disclose its artificial nature in all situations where such interaction is reasonably foreseeable. This effectively shifts the threshold from disclosure only where interaction is certain to wherever interaction is plausible, expanding the obligation for autonomous systems that operate without a clearly defined human counterparty.

Article 50(2): What Must Be Marked and What Falls Outside Scope

Article 50(2) Natural persons must be able to know when they are interacting with AI systems that generate or manipulate synthetic image, video, audio or text content. Article 50(2) imposes a marking and detection obligation for providers.

The Guidelines clarify that Article 50(2) is not limited to GPAI systems but applies broadly to any AI system that generates or manipulates synthetic content. Article 50(2) covers both purpose-built generative systems and multi-purpose or GPAI systems, including agentic AI.

The Guidelines also note that Article 50(2) does not apply where an AI system merely reproduces existing content without generating new or materially altered outputs. This includes systems such as playlists, recommender engines, or data‑recording tools that only observe, record, or relay information without modifying it.

The Guidelines indicate that source code is not within the scope of Article 50(2) because it is treated as a technical output rather than human‑facing content capable of misleading users. However, this exclusion may not be absolute. It likely applies to source code itself, including elements embedded within it such as comments or docstrings, but does not extend to separate natural-language outputs generated by the same system, such as documentation, explanatory text, or user-facing descriptions. Those outputs would fall back within Article 50(2) as standard text capable of being interpreted by users. While the precise boundary remains uncertain, providers should not assume a blanket exemption and should structure transparency measures accordingly.

A seemingly narrow exception for certain industrial or business‑to‑business AI applications is identified by the Guidelines, where marking and detection may not be required. However, this exemption applies only in limited circumstances where outputs are strictly technical (e.g. engineering or internal workflows) and are accessible solely to a restricted group of professionals within an organisation, with safeguards to prevent wider dissemination. This carve-out should be narrowly construed and does not apply where outputs could be shared externally or interpreted by a broader audience.

The Guidelines state that the proposed AI Omnibus package, which remains under consideration by the EU co-legislators, would introduce a limited grandfathering rule for the marking and detection obligations under Article 50(2). That proposed transitional measure would apply only to generative AI systems placed on the market or put into service before 2 August 2026, giving providers additional time to bring those systems into conformity.

The Guidelines clarify that AI-generated or manipulated outputs falling under Article 50(2) already in the information ecosystem prior to 2 August 2026 do not need to be marked or labelled retrospectively. Nevertheless, the Guidelines encourage deployers who possess or disseminate such content to voluntarily apply marking or labelling where appropriate.

Conclusion

In Part 2 of this series, we examine the Commission’s interpretation of Articles 50(3) and 50(4) of the AI Act, and how providers and deployers can show they are complying with their transparency obligations.

Contributed by Isaac Kelly