Stricter rules for General Purpose AI (GPAI), are you prepared for phase two of the AI Act?

The AI Act is being implemented in phases. Since February 2, 2025, the rules around prohibited AI systems and the obligation to work on AI literacy have already applied. The next step will take effect on August 2, 2025. Then new requirements will apply to AI models intended for general applications: the so-called General Purpose AI (GPAI). In this article, we explain what this precisely means.

What is General Purpose AI (GPAI)?

The AI Act categorizes AI systems into four risk levels: prohibited, high risk, low risk, and minimal risk. GPAI models are added as a separate category alongside this classification. The reason for this is that GPAI models are intended for ‘general purposes’ and can be used for various applications. GPAI models serve as the foundation for different AI systems. For those systems, their risk level depends on the specific use.

Precisely because these GPAI models are so versatile, it is difficult to foresee all possible effects and implications of their use. Effective risk management is crucial, which is why additional transparency requirements apply to GPAI model providers. For GPAI models with a so-called system risk (with very high computing power and thus potentially serious consequences for, for example, public health and safety, or possible negative effects on democratic processes), stricter requirements apply.

Examples of GPAI models that are widely used in the cultural sector include:

• Language models such as GPT-4 (on which ChatGPT is built)
• Translation tools such as Google Translate and DeepL
• Image generators such as Stable Diffusion (on which Midjourney is built)
• Programming assistants such as Copilot and Cursos

Transparency obligation

Providers of GPAI models must comply with a number of (transparency) rules:

The EU has also developed a (voluntary) code of conduct to support GPAI model providers in complying with their legal obligations. The code of conduct was drawn up in consultation with a large number of parties – in addition to GPAI model providers and academia, cultural organizations and rights-holder organizations are also represented. More information about the code of conduct can be found here.

Transparency obligations for providers and users of certain AI systems

Transparency obligations also apply to providers and users of certain AI systems. 

For example, the transparency requirements apply to providers of tools involving direct interaction with people, such as a chatbot on your website. You are then required to clearly inform visitors to your site that they are interacting with AI and not with a natural person. It is important to realize that if you modify an existing AI system or tool or develop one yourself, you may fall under the provider rules.

Deepfakes

Special attention is given in the AI Act to AI systems that generate so-called ‘synthetic’ content: deepfakes. The AI Act uses a very broad definition for this: all ‘AI-generated or manipulated image, audio, or video material that resembles existing people, objects, or events, and could be wrongly perceived as authentic by a person.’

Because deepfakes can be harmful (think of deception or fraud), they must always be clearly identifiable as AI content. This can be done via a watermark, metadata, or a digital ‘fingerprint.’ The rules do not apply to content that does not resemble existing people or things – such as cartoons.

There is also an exception for work that is clearly part of an artistic, creative, satirical, fictitious analog work, or program. In principle, a transparency obligation still applies here, but it should not restrict artistic freedom. In such cases, a disclaimer at the start or end of the exhibition, the work, or in the colophon suffices.

News articles and other texts

For AI-generated texts, the rules are slightly different: for texts you publish ‘to inform the public about matters of general interest’ – such as news articles – you must disclose that the text was artificially generated or edited. If you edit or review the text yourself before publication, this obligation does not apply.