Transparency and general-purpose AI models

The four transparency obligations

Article 50 requires that people know they are interacting with an AI system, unless this is obvious (paragraph 1); that synthetic audio, image, video or text outputs carry a machine-readable marking that allows them to be detected (paragraph 2); that people exposed to emotion recognition or biometric categorisation are informed (paragraph 3); and that it is disclosed when a deep fake has been artificially generated or manipulated (paragraph 4).

Text generated to inform the public on matters of public interest must also be labelled, unless it has undergone human review or editorial control with someone holding editorial responsibility (Article 50(4)). The information is given clearly at the latest at the time of the first interaction (Article 50(5)).

Transparency dates

Article 50 applies from 2 August 2026, the general date of application of the Regulation (Article 113). The Omnibus gives generative systems already placed on the market before that date until 2 December 2026 to comply with the marking in paragraph 2 (Article 111(4)).

General-purpose AI models

Their providers must maintain technical documentation of the model, give those who integrate it the information they need to meet their own obligations, have a policy to comply with copyright and publish a sufficiently detailed summary of the training content (Article 53(1)). Models released under a free and open-source licence are exempt from the first two, unless they pose systemic risk (Article 53(2)).

Systemic risk is presumed when training exceeds 10^25 floating-point operations (Article 51(2)), and then evaluations with adversarial testing, systemic risk management, incident reporting and cybersecurity are added (Article 55).

These rules have applied since 2 August 2025 (Article 113, point (b)); models placed on the market earlier had to be brought into line by 2 August 2027 at the latest (Article 111(3)). A company that uses a third party's model through an application is not a provider of the model: where applicable, it answers as provider or deployer of the system it builds.

What an auditor usually asks for

• Chat assistants identified as AI to customers and employees
• Synthetic content published with a notice where there is no editorial review
• General-purpose model documentation requested from the provider when it is integrated into your own system
• Use of emotion recognition or biometric categorisation disclosed to the people exposed

How to tackle it in practice

1. 1Review every channel where a customer or applicant might converse with an AI and add the notice wherever it is not obvious.
2. 2Decide who signs off as editorially responsible for texts published with the help of AI.
3. 3Ask the model provider for the information in Article 53(1)(b) before building your own system on top of it.

Where each figure comes from

References to the official text in force. If a standard is revised, this page is updated and the review date says so.

1. Regulation (EU) 2024/1689 laying down harmonised rules on artificial intelligence (Artificial Intelligence Act) EUR-Lex · OJ L 2024/1689 · 12 July 2024
2. Regulation (EU) 2026/1744 (Digital Omnibus on AI), amending Regulation (EU) 2024/1689 EUR-Lex · OJ L 2026/1744 · 24 July 2026
3. Regulation (EU) 2024/1689, consolidated text EUR-Lex · CELEX 02024R1689-20260727 · Version of 27 July 2026