fdg

Join us as we uncover the latest insights into the European Union’s (the ‘EU’) ground-breaking Artificial Intelligence Act, in the wake of the leak of its most recent consolidated version on 26 January 2024 [i]. We delve herein into the core objectives, risk-based categorisation, and key provisions of this text, originally proposed by the European Commission on 21 April 2021 and now nearing adoption (the ‘AI Act’).

Starting with its primary objective, the AI Act aims to enhance the functioning of the internal market by establishing a uniform legal framework for the development, marketing, and use of artificial intelligence systems (the ‘AI systems’) within the EU. This framework is to be applied in accordance with EU values, to ensure a high level of protection of health, safety and fundamental rights enshrined in the Charter of Fundamental Rights of the EU.

The AI Act introduces a risk-based approach for AI systems, similar to the GDPR’s data protection rules. This approach distinguishes three AI systems, based on the risk posed by their usage: Low or Minimal Risk AI Systems, High-Risk AI Systems, and Prohibited (or unacceptable) AI Systems. Occasionally, a fourth level is also mentioned in some papers published online: the ‘transparency risk’ or ‘limited risk’.

This risk differentiation stands out from Recital (14) of the AI Act, which currently states the following:

“It is therefore necessary to prohibit certain unacceptable artificial intelligence practices, to lay down requirements for high-risk AI systems and obligations for the relevant operators, and to lay down transparency obligations for certain AI systems”.

Below is a table detailing the three risk levels of AI systems, along with examples of AI systems for each risk level, and listing some requirements and obligations associated with each of those levels.

The AI Act also briefly touches upon the so-called general-purpose AI systems.

These AI systems are based on or integrate a general-purpose AI model with the capability of serving a variety of purposes (Article 3(44e) AI Act; Recital (60d) AI Act). However, the AI Act contains no specific requirement or obligation for general-purpose AI systems. Instead, it focuses on general-purpose AI models, which serve as a foundation of these general-purpose AI systems. This may include some of OpenAI’s ChatGPT models.

A general-purpose AI model may be placed on the market in various ways, including through libraries, application programming interfaces (APIs) or as a direct download (Recital (60a) AI Act). They allow for flexible generation of content (such as text, audio, images, or video) and can readily accommodate a wide range of tasks (Recital (60c) AI Act).

General-purpose AI models are essential components of some AI systems, but they do not constitute AI systems on their own – notably because they lack further components to be considered as such, for example, a user interface (Recital (60a) AI Act). Therefore, contrary to what some papers may suggest, they are not inherently an AI system, nor do they fall in one of the existing three risk categories of AI systems.