Franchetto Legal
Digital legal practice · data law
AI Act · Article 50 · transparency

The AI Act's transparency obligations (Article 50): what applies from 2 August 2026

From 2 August 2026 the transparency obligations of Article 50 of the AI Act apply. There is a single calendar exception, and it concerns the marking of synthetic content.

In force since 2 August 2026 · next deadline 2 December 2026

In short

From 2 August 2026, the transparency obligations of Article 50 of the AI Act (Regulation (EU) 2024/1689) apply: anyone who provides or uses certain artificial-intelligence systems must disclose it to people. There is a single calendar exception: for the marking of synthetic content (Article 50(2)), systems already on the market before 2 August 2026 have until 2 December 2026, which is now the next deadline to watch. This is set by the Digital Omnibus on AI (Regulation (EU) 2026/1744, in force since 27 July 2026). Penalties reach up to EUR 15 million or 3% of total worldwide annual turnover.

The four transparency obligations of Article 50

Article 50 is not a single obligation: it imposes transparency in four distinct situations. Two of the four obligations fall on whoever develops and supplies the system (the provider), two on whoever uses it (the deployer).

  • Human-machine interaction (Article 50(1)), a provider obligation. Systems intended to interact with people (chatbots, voice assistants) must be designed so the user knows they are dealing with an artificial intelligence, unless it is already obvious.
  • Synthetic content (Article 50(2)), a provider obligation. Systems that generate artificial audio, images, video or text must mark the output as AI-generated, in a machine-readable and detectable format.
  • Emotion recognition and biometric categorisation (Article 50(3)), a deployer obligation. Those who use these systems must inform the people exposed to them.
  • Deepfakes and public-interest text (Article 50(4)), a deployer obligation. Manipulated image, audio or video content (deepfakes) must be declared as artificial or manipulated; the same applies to text published to inform the public on matters of public interest. The obligation is subject to some exceptions (see the FAQ).

Provider

whoever develops or places the system on the market

  • Article 50(1)

    Human-machine interaction

    Chatbots and voice assistants must be designed so the user knows they are dealing with an AI, unless it is already obvious.

  • Article 50(2)

    Synthetic content

    Artificial audio, images, video or text must be marked as AI-generated, in a machine-readable and detectable format.

Deployer

whoever uses the system in their own business

  • Article 50(3)

    Emotions and biometrics

    Those who use emotion-recognition or biometric-categorisation systems must inform the people exposed to them.

  • Article 50(4)

    Deepfakes and public-interest text

    Manipulated content (deepfakes) and public-interest text must be declared as artificial or manipulated.

Summary of the obligations and deadlines; for exceptions and details, see the article text.

The picture became clearer just before the obligations applied: on 20 July 2026 the European Commission published its guidelines on the implementation of the Article 50 transparency obligations, following a public consultation. They are not law: they are non-binding interpretive guidance. They do, however, help clarify who is covered and how to handle interactive systems, the labelling of AI-generated content and deepfakes. The technical marking solutions (metadata, watermarks) are instead detailed in the accompanying Code of Practice on the transparency of AI-generated content.

The two dates that matter

2 August 2026: the base date. The Article 50 transparency obligations are applicable as of this date (Article 113 of the AI Act). The Digital Omnibus did not postpone this deadline.

2 December 2026: the only transitional exception. It concerns exclusively the marking of synthetic content (Article 50(2)) and only systems placed on the market before 2 August 2026. According to the European Commission, the Digital Omnibus (Regulation (EU) 2026/1744) postponed to 2 December 2026 the marking of synthetic content for systems placed on the market before 2 August 2026.

  1. 2 Aug 2026

    The base date

    The Article 50 transparency obligations become applicable (Article 113). The Digital Omnibus did not postpone this deadline.

  2. 2 Dec 2026Exception

    The only transitional exception

    Only the marking of synthetic content (Article 50(2)), and only for systems placed on the market before 2 August 2026 (Digital Omnibus, Reg. (EU) 2026/1744).

Summary of the obligations and deadlines; for exceptions and details, see the article text.

Two points the extension does not cover:

  • Systems placed on the market after 2 August 2026 must comply with Article 50(2) immediately, that is from 2 August: no extension.
  • All the other Article 50 obligations (paragraph 1 interaction, paragraph 3 biometrics and emotions, paragraph 4 deepfakes and public-interest text) remain fixed at 2 August 2026.

Provider or deployer: who must do what

The practical question is not «am I subject to Article 50?» but «in which role?».

  • If you develop or place on the market a system (a chatbot, an image or text generator), the obligations that concern you are the provider's: disclose the AI interaction (paragraph 1) and mark synthetic content (paragraph 2).
  • If you use an AI system in your business (emotion recognition, biometric categorisation, production of deepfakes or informational text), the obligations are the deployer's: inform the people exposed (paragraph 3) and declare manipulated content (paragraph 4).

Many organisations are both across different systems: provider of their own chatbot, deployer of a third-party tool. The mapping must be done system by system.

The penalties

Breaching the Article 50 transparency obligations is subject to the AI Act's penalty regime (Article 99): up to EUR 15 million or, if higher, up to 3% of total worldwide annual turnover for the preceding financial year.

For SMEs, including start-ups, the ceiling is calculated downwards: the lower of the fixed amount or the percentage applies (Article 99(6)). In practice, for a smaller business the real cap is almost always the 3% of turnover, not the EUR 15 million: the higher headline figure is theoretical, and what bites is the percentage of turnover. Beyond the amount, the risk is twofold: the fine and the reputational damage of AI use perceived as insufficiently transparent towards clients and users.

What to do now: the checklist

Five concrete steps:

  1. Inventory your AI systems. Map where you use AI in contact with people or with their content: chatbots and assistants, text or image generators, behavioural analysis, editing tools.
  2. Classify each use against Article 50. For each system, establish which of the four obligations applies and whether you are a provider or a deployer (often you are the deployer of a third-party model).
  3. Implement the disclosure and the labelling. A «you are interacting with a virtual assistant» notice for chatbots (paragraph 1); an AI-generated content label for images, video and text (paragraph 2); a notice to exposed persons for emotion recognition and biometrics (paragraph 3). Per the Commission guidelines, disclosure may be omitted only where it is obvious, to a reasonably well-informed, observant and circumspect person, that this is an AI: this exception is to be interpreted restrictively, and it is for you to justify relying on it. The information must be provided in a clear and distinguishable manner, at the latest at the time of the first interaction or exposure (Article 50(5)): it cannot simply be buried in the terms and conditions.
  4. Synthetic content: verify the technical marking. Article 50(2) requires a machine-readable and detectable format: a visible caption is not enough. The marking must be effective, reliable and aligned with the state of the art, for example through metadata or watermarks, as specified by the Code of Practice on the transparency of AI-generated content. Ask the model provider which technique it uses.
  5. Document. Keep a record of the classification and the measures adopted: in the event of a check, the evidence of what you did is what counts. Transparency must be demonstrated, not merely asserted.

Frequently asked questions

Did the AI Act postpone the transparency obligations?

No. The Article 50 obligations apply from 2 August 2026. The Digital Omnibus (Regulation (EU) 2026/1744) provided a single transitional exception: the marking of synthetic content (Article 50(2)) for systems placed on the market before 2 August 2026, with a deadline of 2 December 2026.

What is the Digital Omnibus on AI?

It is Regulation (EU) 2026/1744, in force since 27 July 2026, which simplifies the implementation of the AI Act. Among other things, it introduced the transitional provision on the marking of synthetic content. It is a separate act from the so-called Data Omnibus (which concerns the GDPR, ePrivacy and other digital rules).

Are the Commission's guidelines on Article 50 binding?

No. The guidelines on the implementation of the Article 50 transparency obligations, published by the European Commission on 20 July 2026 after a public consultation, are non-binding interpretive guidance. They help clarify how to apply the obligations; the technical detail on marking synthetic content sits in the accompanying Code of Practice.

Does the extension to 2 December 2026 apply to all AI-generated content?

No. It applies only to systems placed on the market before 2 August 2026. New systems, placed after 2 August, must comply with Article 50(2) from 2 August 2026.

Who must mark synthetic content, the provider or the user?

The machine-readable marking of AI-generated content (Article 50(2)) is a provider obligation. The declaration of deepfakes (Article 50(4)) is instead a deployer obligation.

Does the synthetic-content marking apply to editing tools too?

Not in every case. The Article 50(2) marking obligation does not apply where the AI system performs an assistive function for standard editing or does not substantially alter the input data provided by the deployer or its semantics. A tool that only makes standard adjustments does not, in itself, trigger the marking obligation.

Are there exceptions to the deepfake and text disclosure obligation?

Yes. For deepfakes forming part of manifestly artistic, creative or satirical works, disclosure is due but in a form that does not hamper the enjoyment of the work. For text published on matters of public interest, the obligation does not apply where the content underwent human review or editorial control and a natural or legal person holds editorial responsibility (an exemption that is relevant for publishers and media outlets).

What are the penalties?

Up to EUR 15 million or 3% of total worldwide annual turnover, under the regime of Article 99 of the AI Act. For SMEs and start-ups, the lower of the two applies (Article 99(6)).

Sources

Leggi questa pagina in italiano