Who is responsible for AI Act transparency obligations?

Michał Rutkowski • for boards, compliance teams and DPOs • published August 21, 2026 • updated October 20, 2026 • 8 min read

Let’s talk

Wondering how this plays out in your organisation? Get in touch — I answer personally.

Short answer

Responsibility follows the role held towards a specific system, not the level of its risk. The provider is responsible for making the system disclose that a person is interacting with AI and for marking synthetic content in a machine-readable format. An organisation that itself develops an AI system or has it developed and puts it into service under its own name is also a provider, including for its own use. The deployer has two obligations of its own. It informs people about emotion recognition and biometric categorisation. It also discloses deep fakes and text published with the purpose of informing the public on matters of public interest. The obligation concerning text does not apply where the content has undergone human review or editorial control and a natural or legal person holds editorial responsibility for the publication. The obligations have applied since 2 August 2026. Providers of systems generating synthetic content that were placed on the market before that date have until 2 December 2026 for marking.

Why this question arises at all

In the summer of 2026 the market settled on one message. The AI Act has been postponed. The postponement, however, covered the requirements and obligations concerning high-risk systems. The transparency provisions became applicable on the original schedule. The exception is a transitional period until 2 December 2026 for providers of systems generating synthetic content that were placed on the market before 2 August 2026. It concerns only the marking and detectability of outputs under Article 50(2).

For an organisation that only uses AI, the difference is fundamental. High-risk requirements apply to a narrow group of use cases. They apply from 2 December 2027 to high-risk AI systems under Annex III and from 2 August 2028 to high-risk AI systems under Annex I. Transparency obligations do not ask about the risk category. They ask what the system does and who uses it in a given arrangement.

The second source of confusion is commercial. The vendor of a tool usually declares compliance with the AI Act, and the organisation reads that as the end of the subject. Yet two of the four transparency obligations are addressed directly to the deployer. The provider cannot take them over from the client, because it does not decide towards whom and for what purpose the client uses the system.

The third source is terminological. “Transparency” in the AI Act does not mean documentation or a policy. It means specific information given to a specific person at a specific moment.

What to establish before deciding

Four obligations and two roles

Article 50 of the Regulation contains four separate obligations. The first two are designed into the product; the other two are performed inside the organisation.

ObligationWho holds itWhat it coversExceptions and limits
Informing people that they interact with an AI system (para. 1) — from 2 August 2026providersystems intended to interact directly with natural personsno notice is required where the AI interaction is obvious to a reasonably well-informed, observant and circumspect person in the circumstances and context of use; a separate law-enforcement exception applies, except for systems available to the public to report a criminal offence
Marking synthetic content in a machine-readable format (para. 2) — from 2 August 2026provideraudio, image, video and text that is generated or manipulatedexcluded for assistive standard editing or where the system does not substantially alter the input data or its semantics; a separate law-enforcement exception applies; providers of systems placed on the market before 2 August 2026 have until 2 December 2026 to comply (Article 111(4))
Informing people about emotion recognition and biometric categorisation (para. 3) — from 2 August 2026deployerthe persons the system is applied tolaw-enforcement exception, subject to safeguards for the rights and freedoms of third parties
Disclosing deep fakes and published text (para. 4) — from 2 August 2026deployerimage, audio and video content constituting a deep fake; text published with the purpose of informing the public on matters of public interestdeep fake in an evidently artistic, creative, satirical or fictional work or in an analogous one — the duty is limited to disclosure in a way that does not hamper the display or enjoyment of the work; text — no obligation after human review or editorial control where a natural or legal person holds editorial responsibility for the publication; a separate law-enforcement exception applies

The split has a practical consequence for procurement. The provider’s obligations can be moved into contractual terms and verified before deployment. The deployer’s obligations cannot be moved anywhere. They depend on how the organisation uses the system’s output, and the provider does not control that.

The role follows from the definition, not from who sold the tool. A provider is an organisation that develops an AI system or has it developed and places it on the market or puts it into service under its own name or trademark (Article 3(3)). This also covers a system built on another company’s model and used only within the organisation (Article 3(11) and (68)). In its guidelines on Article 50 the European Commission gives the example of a chatbot developed in-house. Such an organisation is responsible for the obligation under paragraph 1 and, where it generates content, also under paragraph 2. According to the Commission, a company commissioning an advertisement without deciding on the agency’s use of AI is not a deployer.

What the deployer has to disclose

The first obligation covers two kinds of systems. The organisation informs people that an emotion recognition system or a biometric categorisation system is being applied to them. The provision also requires personal data to be processed in line with the GDPR, so both layers run in parallel.

In the workplace, however, the question usually comes earlier. The Regulation prohibits the use of AI systems to infer emotions of individuals in the workplace and in education institutions. The exception covers medical and safety reasons only. Before the organisation designs information for its staff, it therefore has to settle whether it may run such a tool at all. For biometric categorisation a separate prohibition under Article 5(1)(g) applies regardless of the setting. With narrow exceptions it covers inferring race, political opinions or sexual orientation from biometric data, among other things.

The second obligation covers content. The Regulation defines a deep fake as AI-generated or manipulated image, audio or video content. Such content resembles existing persons or objects, places, entities or events. A person could falsely take such content to be authentic or truthful. The European Commission’s guidelines on Article 50 read the word “existing” broadly. According to the Commission, a realistic synthetic avatar of a person who could plausibly exist can also be a deep fake. The assessment of whether content could appear authentic is objective in the Commission’s view. It does not require any intention to deceive. For such content the deployer discloses that it has been artificially generated or manipulated.

The same paragraph covers text. The obligation for text is narrower, however, and concerns text published in order to inform the public on matters of public interest. Ordinary marketing communication and internal correspondence fall outside its scope. The line gets thin in the communication of public institutions and in materials on safety, health and the environment.

When the editorial exemption applies

For text the provision sets out an exemption that is easy to apply too broadly. The duty to disclose falls away where the content has undergone human review or editorial control and a natural or legal person holds editorial responsibility for the publication.

The condition has two limbs. Review alone is not enough — there has to be someone bearing editorial responsibility. In its guidelines on Article 50 the European Commission states that the review must concern the substance, with fact-checking as a minimum. According to the Commission, neither spell-checking nor cursory approval nor the mere existence of an editorial policy meets that condition. The Commission also considers that a substantive AI change to the text after approval voids the exemption. It recommends making the details of the person or function with editorial responsibility publicly available, for example in the website’s legal notice. An organisation relying on the exemption would therefore do well to document who reviews the text before publication and how.

A separate relaxation applies to deep fakes forming part of an evidently artistic, creative, satirical or fictional work or of an analogous one. The obligation is then limited to disclosing the existence of such content in a way that does not hamper the display or enjoyment of the work.

How and when to give the information

The Regulation settles both the form and the moment. The information is given in a clear and distinguishable manner, at the latest at the time of the first interaction or exposure. For content, the Commission takes this to mean the moment a person is first exposed to the system’s output. The information must also meet the applicable accessibility requirements.

Three practical design decisions follow from that. The information cannot be hidden in terms of service or in a privacy policy, because it does not reach the person at the moment of first contact. It must also be accessible to persons with disabilities, because the Commission treats this as a condition of clear information. And it cannot be delivered after the fact.

The provision also states that paragraphs 1 to 4 are without prejudice to other transparency obligations of deployers under Union or national law. Information under Article 50 therefore does not replace the information duty under the GDPR. They are two parallel messages with different subject matter.

What changes on 2 December 2026

The July 2026 omnibus added one date that concerns older tools. Providers of systems generating synthetic content that were placed on the market before 2 August 2026 have until 2 December 2026 to align with the marking obligation.

For the deploying organisation the conclusion runs the other way than it first seems. A tool generating content that was placed on the market before 2 August 2026 may not mark its output in a machine-readable way until December. What counts is the date on which the provider placed it on the market, not the date of purchase. The disclosure obligations under paragraph 4, however, already rest with the deployer and apply regardless of what the tool does. In that period the disclosure has to be delivered by a process, not by a product feature.

What you can rely on

On 20 July 2026 the European Commission published guidelines on Article 50. It approved their content at that time and announced formal adoption once all language versions are available. The guidelines explain the division of responsibility between the provider and the deployer, the scope of the exemptions and the meaning of a deep fake, among other things. They are not binding, but they show how the Commission reads the provision.

The second document is the Code of Practice on Transparency of AI-Generated Content. Article 50(7) provides that the Commission assesses whether adherence to such a code is sufficient to ensure compliance with paragraphs 2 and 4. In doing so the Commission takes utmost account of the opinion of the Board. The second section of the code is addressed directly to deployers and describes the labelling of deep fakes and of text under paragraph 4 as well as internal processes. According to the Commission’s guidelines, adhering to a code assessed as adequate makes it possible to demonstrate compliance with those paragraphs. Signing up to the code is therefore a way of demonstrating compliance, not compliance itself.

Most common mistakes

  • Assuming the postponement covered the whole Regulation. What moved were the requirements for high-risk systems. Transparency has applied since 2 August 2026. The transitional period until 2 December 2026 covers only the marking of outputs of older systems generating synthetic content.
  • Assigning all four obligations to the provider. Paragraphs 3 and 4 of Article 50 are addressed to the deployer and cannot be shifted by contract.
  • Treating machine-readable marking as disclosure to the recipient. The marking under paragraph 2 serves technical detectability. The disclosure under paragraph 4 is a message for a human being.
  • Invoking the editorial exemption with no one holding editorial responsibility. The provision requires both elements at once, and the organisation has to be able to show both.
  • Putting the information in the terms of service instead of at the moment of first contact. The provision fixes a moment, not a place in the documentation.
  • Overlooking the prohibition on emotion recognition in the workplace. Informing staff does not cure a practice that is prohibited.

Conclusions and next steps

Transparency, alongside AI literacy and prohibited practices, already concerns an organisation using off-the-shelf tools.

Remember

Responsibility for transparency is decided by the role held towards a specific use case, not by the risk category and not by the provider’s declaration.

The organisation itself is responsible for two obligations, so it needs a process rather than just a clause in a contract.

A sequence that works when putting this area in order.

  1. Collect the use cases in which AI interacts with people or produces synthetic content. Add the emotion recognition and biometric categorisation systems applied to people.
  2. For each use case establish the organisation’s role and assign the obligations from the relevant paragraph of Article 50.
  3. Check whether tools placed on the market before 2 August 2026 mark their output in a machine-readable way. Plan a solution for the period until December.
  4. Design the wording and the moment of the information, and for text covered by Article 50(4) settle who holds editorial responsibility.
  5. Keep a trace of those decisions, because that is what answers the question of a client, an auditor or an authority.

A trace of the decisions is easy to skip. Transparency obligations are performed towards people, but they are demonstrated towards an authority — and that calls for different material than a notice on a website.

Sources

  • Regulation (EU) 2024/1689 of the European Parliament and of the Council (Artificial Intelligence Act), consolidated version of 27 July 2026, Article 3(3), (11), (60) and (68), Article 5(1)(f) and (g), Article 25, Article 50, Article 111 and Article 113 — eur-lex.europa.eu.
  • Regulation (EU) 2026/1744 of the European Parliament and of the Council of 8 July 2026, Article 1(39) and (40) (amendments to Articles 111 and 113) — eur-lex.europa.eu.
  • European Commission, guidelines on the implementation of the transparency obligations for certain AI systems under Article 50 of Regulation (EU) 2024/1689 — content approved by Communication C(2026) 5054 final of 20 July 2026, formal adoption announced once all language versions are available — digital-strategy.ec.europa.eu.
  • Code of Practice on Transparency of AI-Generated Content, together with the Commission’s opinion on its assessment, European Commission — digital-strategy.ec.europa.eu.

See where AI already speaks for your organisation

Transparency obligations concern use cases that usually appeared without a board decision. The review starts with a list of them and ends by assigning the obligations under Article 50 to specific people in the organisation.

Author

Michał Rutkowski — LabLogic. He combines the legal, organisational and technological perspective in his work with the boards of medium-sized and large organisations: support for and performance of the DPO role, preparation for NIS2 and the Polish KSC Act, incident response, audits and training. Contact: M.Rutkowski@LabLogic.pl · LinkedIn

This material is general and educational in nature. It is not an individual legal opinion or a recommendation for any specific organisation. The scope of the obligations should be assessed against the situation of the organisation concerned.

Legal status: November 2026.

Scroll to Top