knowledge 19 August 2026 |

AI Transparency: European Commission’s Guidelines on Article 50 – Part 1 (Provider Obligations)

On 20 July 2026, the European Commission (the Commission) published its finalised, non-binding Guidelines on the transparency obligations under Article 50 (the Guidelines), which aim to reduce the risks of impersonation, deception, misinformation, manipulation and fraud arising from AI systems.

This Part 1 briefing seeks to provide an overview of the Guidelines relating to the obligations on providers of AI systems under Articles 50(1) and 50(2) of the AI Act.

By way of reminder, a provider is an organisation that develops an AI system (or has one developed for it) and then makes that system available to others – whether by selling it, licensing it, or putting it into use – under its own name or brand.

Obligations on Providers of AI Systems

Article 50(1) – Transparency for Interactive AI systems

Article 50(1) applies to AI systems intended to interact directly with natural persons e.g. customer service chatbots, AI assistants, AI companions and AI bots on social media. Providers must design these systems so that a natural person is made aware, no later than upon the first interaction, that they are interacting with an AI system, with the notification embedded in the system’s design and development.

No particular disclosure format is prescribed: textual, auditory, visual or multimodal techniques are all acceptable, provided persons are properly and effectively informed without risk of deception or manipulation. Disclosures should be adapted for vulnerable users, including children, older persons, persons with disabilities, with standardised practices and labels recommended for consistency and user awareness, without disrupting the user experience or encouraging “banner blindness” (whereby users are inclined or habitualised into ignoring website banner information).

Disclosures should be proportionate. A single prominent notification before first interaction will usually suffice, but riskier contexts (such as interaction with vulnerable users, or the giving of health advice) may warrant periodic reminders. Disclosures hidden in terms and conditions, imperceptible machine-readable markings, or ambiguous signals are insufficient; the system must be designed to disclose its nature whenever asked, or where it can reasonably be assumed that the person is likely to be misled or confused about the AI origin of the interaction.

Two exemptions apply to the transparency obligations under Article 50(1): first, where the AI nature of the interaction is obvious to a reasonably well-informed, observant and circumspect user (an exception to be construed narrowly); and second, where the system is authorised by law to detect, prevent, investigate or prosecute criminal offences, subject to appropriate safeguards for third parties’ rights and freedoms.

Article 50(2) – Marking and Detection of AI-generated or Manipulated Content

Article 50(2) requires providers of AI systems that generate or manipulate synthetic content to ensure outputs are (i) marked in a machine-readable format and (ii) capable of being detected as AI-generated or AI-manipulated. The Guidelines make clear both elements must be satisfied.

This applies to synthetic audio, image, video or text outputs including virtual and augmented reality content and content mixed with human-created material, but excludes standard editing (e.g. spell-checking, grammar correction, basic translation, formatting changes and noise removal) and content based on simple data processing, mere reproduction of existing content (e.g. playlists), unaltered recordings, short outputs (e.g. single words or user interface labels), and source code or configuration files. The obligation may still apply to AI agents whose outputs are AI-generated or manipulated content perceptible by natural persons, but not to intermediate processing steps or non-user-facing outputs. Ephemeral, unrecorded real-time content (such as video games) may also be exempt from marking where technically infeasible and users are otherwise made aware of its artificial origin, as is content that is generated or manipulated for law enforcement purposes as authorised by law.

Providers may use a single technique or combination (e.g. watermarking, metadata, cryptographic methods) at system or model level to mark materials as AI-generated, and may rely on solutions from upstream providers or third parties, though responsibility for compliance remains with the provider.

Detection tools must be available to persons likely to encounter the content, produce human-readable results, and be interoperable with other detection systems. All technical solutions must be effective, interoperable, robust and reliable, assessed against what is technically feasible and the state of the art, with providers expected to update their solutions as standards evolve.

Separately, actors that merely disseminate or transmit AI-generated content without exercising authority over the system, such as hosting providers, online platforms and broadcasters, are not considered “deployers”, though they are encouraged to preserve labelling and help users detect artificial content.

In addition, the Guidelines clarify that Article 50 does not explicitly apply to GPAI models, though transparency measures may be implemented at model level where an in-scope system is built on a GPAI model from the same provider, and other GPAI model providers are encouraged to do so voluntarily.

What Else?

Part 2 in this series covers the Guidelines relating to the obligations on deployers of AI systems under Articles 50(3) and 50(4), the general requirements for transparency disclosures under Article 50(5), and the related Code of Practice and enforcement regime. It also sets out some suggestions of what organisations navigating this regulatory landscape should do next to achieve compliance.

How Can McCann FitzGerald LLP Help?

For further information or assistance, please reach out to one of the key contacts below, or your usual contact at McCann FitzGerald LLP.

Also contributed to by Eva Keane.