Article 50 AI Act: Transparency Rules Apply 2 August 2026

On 20 July 2026 the European Commission published its guidelines on Article 50 of the AI Act, the provision that governs transparency for AI systems. Less than two weeks later, on 2 August, those obligations become applicable to anyone who places on the market or puts into service in the Union a system that interacts with people or generates synthetic content.

The guidelines are non-binding, and any authoritative interpretation will ultimately rest with the Court of Justice, yet they set a direction that is hard to misread. Transparency belongs to the architecture of the system and to the entire chain that produces and distributes the content. A label added at the end of the line arrives too late. Hiding in the terms of use that "the service uses AI" is not enough, and the guidelines say so plainly: a disclosure contained only in contractual terms, a URL, or technical documentation does not satisfy the obligation.

Who marks and who discloses

Article 50 splits four obligations between two figures. The provider, whoever develops the system and places it on the market under their own name, must design interactive systems so that a person knows they are talking to an AI (paragraph 1) and mark in a machine-readable format the synthetic content the system generates or manipulates (paragraph 2). The deployer, whoever uses the system under their own authority for professional purposes, must inform those exposed to emotion recognition or biometric categorisation systems (paragraph 3) and visibly disclose deep fakes and public-interest text produced or manipulated by AI (paragraph 4).

The distinction carries a practical consequence many organisations underestimate. The provider's technical marking, a watermark or a piece of metadata, never reaches the eye of the person looking at the content, and the Commission clarifies that the deployer cannot lean on that marking to satisfy the labelling duty under paragraph 4. Two separate layers are needed: the machine-readable signal upstream and the human-readable disclosure downstream, and they answer to different actors.

AI agents disclose on whose behalf they act

Among the densest passages is the treatment of agents. An AI agent that books, handles correspondence, or concludes contracts on someone's behalf falls under paragraph 1, and must disclose two things at once: its artificial nature and the person on whose behalf it is acting. To transparency of identity the guidelines add transparency of delegation, and with it the question of who answers for the consequences.

When the provider cannot know in advance whether and with whom the agent will interact once released, the guidelines ask for the disclosure to be designed at the architecture level, so that the agent reveals itself whenever an encounter with a person is reasonably likely. It must also identify itself to the person instructing it, at key moments such as authorising or validating an action, and at every new interaction. This is a reading that looks ahead, to the multi-agent architectures where one agent talks to another before reaching the human being.

Editorial control that counts, and the kind that does not

AI-generated or manipulated text published to inform the public on matters of public interest must be labelled as artificial. There is an exception, and here the Commission raises the bar. The exemption applies only if the text has undergone human review or editorial control that engages with the substance, and only if a person, natural or legal, holds editorial responsibility for it.

Substantive is the word that carries the weight. Spell-checking, a purely formal review, or the mere existence on paper of an editorial policy does not trigger the exemption. The review has to reach the content, with fact-checking as the minimum, and whoever performs it must have the authority to approve, alter, or reject on the merits. There is more: if AI intervenes substantively after editorial sign-off, the exception falls away and the text becomes labellable again. For newsrooms and for anyone producing content at scale, this is an operational discipline before it is a legal rule.

The window closes on 2 August

From 2 August 2026 every system within the scope of Article 50 placed on the Union market must be compliant, regardless of when it was put into service. The AI Omnibus, recently adopted by the Union legislature, opens a single window of tolerance: generative systems already on the market before 2 August have until 2 December 2026 to bring their marking and detectability under paragraph 2 into conformity. The interaction disclosure of paragraph 1 and the deployer's labelling under paragraph 4 apply from 2 August instead, with no extension.

The penalty sets the stakes. Anyone breaching the transparency obligations risks up to 15 million euro or, for an undertaking, up to 3% of total worldwide annual turnover, whichever is higher. For SMEs and start-ups the lower ceiling applies. The Commission also points to a route for demonstrating compliance: adhering to the Code of Practice on Transparency of AI-Generated Content, published on 10 June 2026, which for signatories focuses supervision on verifying adherence.

The numbers show how real the problem is. In Italy 71% of large companies run at least one active AI project, while only 9% have structured governance (Osservatorio Artificial Intelligence, Politecnico di Milano, February 2026). Article 50 requires knowing, for each system, whether it generates content, whether it interacts with people, and who answers for disclosure. That is a governance question before it is a compliance one.

Knowing when there is no person behind a text

The purpose of these obligations is set out in Recitals 132 and following: to reduce the risks of impersonation and manipulation at scale, and to protect trust in the information ecosystem. The guidelines try to preserve, in a communication space mediated by AI, the minimum conditions for a person to choose whether and how much to trust a voice, a face, a text. Knowing when there is no human being behind a piece of content becomes a precondition of autonomy, not a compliance detail.

For a company that designs or adopts AI systems, the centre of gravity shifts. Transparency stops being a box to tick just before launch and becomes a design constraint, to be decided before building. Whoever reaches 2 August having already mapped their systems and the related disclosure responsibilities will not be playing catch-up. A few weeks remain to work out, system by system, which side of the line they fall on.

Checklist: what to put in order before 2 August

• Inventory and classification. Map every AI system in use or coming, and classify it: does it interact with people, generate or manipulate content (image, audio, video, text), recognise emotions, categorise biometric data.
• Role per system. Establish whether you are the provider, the deployer, or both. Obligations differ: technical marking sits with the provider, perceivable labelling with the deployer.
• Interactive systems. Check that chatbots, voice assistants, and agents disclose their artificial nature at first contact, clearly and distinguishably, not only in the terms of use or documentation.
• AI agents. Make sure they disclose both their artificial nature and on whose behalf they act, at the architecture level, and to whoever instructs them at the points where they authorise or validate an action.
• Synthetic content (provider). Implement machine-readable marking and detection that are effective and interoperable. Generative systems already on the market before 2 August have until 2 December 2026 for this obligation only.
• Deep fakes and public-interest text (deployer). Apply a label perceivable to the human eye. The provider's technical marking does not cover this obligation.
• Emotion recognition and biometric categorisation (deployer). Inform the people exposed to the system.
• Editorial exception. If you publish public-interest text without a label, document substantive human review (with fact-checking) and make editorial responsibility publicly identifiable.
• Accessibility and vulnerable groups. Make the disclosure accessible and suitable for minors and vulnerable people when they are reasonably part of the audience.
• Proof of compliance and internal ownership. Consider adhering to the Code of Practice on Transparency of AI-Generated Content, coordinate disclosure with GDPR and the DSA where they apply, and assign who answers for disclosure on each system.

---

At ZeroFive we work with CEOs and boards to map the AI systems in use and decide where transparency has to be designed in, before it becomes a compliance problem. The starting point is the AI Rating. To talk it through: calendly.com/fabiolalli/zerofive.

Where Art. 50 sits in the governance picture

Article 50 transparency is not an isolated task: it is one control inside a wider system. Companies that already run AI governance consulting handle it as a design requirement, the others meet it as an emergency. For full alignment with the regulatory deadlines the track is EU AI Act compliance consulting, and the service overview is on AI consulting for enterprises.