Artificial Intelligence Act

The Artificial Intelligence Act (AI Act) is a European Union regulation concerning artificial intelligence (AI). It establishes a common regulatory and legal framework for AI within the European Union (EU). The regulation entered into force on 1 August 2024, with provisions that shall come into operation gradually over the following 6 to 36 months.

It covers most AI systems across a wide range of sectors, with exemptions for AI used only for military, national security, research purposes, or for non-professional use. As a form of product regulation, it does not create individual rights; instead, it places duties on AI providers and on organisations that use AI in a professional context.

The Act classifies non-exempt AI applications by their risk of causing harm. There are four levels – unacceptable, high, limited, minimal – plus an additional category for general-purpose AI.

Applications with unacceptable risks are banned.

High-risk applications must comply with security, transparency and quality obligations, and undergo conformity assessments.

Limited-risk applications only have transparency obligations.

Minimal-risk applications are not regulated.

For general-purpose AI, transparency requirements are imposed, with reduced requirements for open source models, and additional evaluations for high-capability models.

The Act also creates a European Artificial Intelligence Board to promote national cooperation and ensure compliance with the regulation. Like the EU's General Data Protection Regulation, the Act can apply extraterritorially to providers from outside the EU if they have users within the EU.

Proposed by the European Commission on 21 April 2021, it passed the European Parliament on 13 March 2024, and was unanimously approved by the EU Council on 21 May 2024. The draft Act was revised to address the rise in popularity of generative artificial intelligence systems, such as ChatGPT, whose general-purpose capabilities did not fit the main framework.

Provisions

Risk categories

There are different risk categories depending on the type of application, with a specific category dedicated to general-purpose generative AI:

Unacceptable risk – AI applications in this category are banned, except for specific exemptions. When no exemption applies, this includes AI applications that manipulate human behaviour, those that use real-time remote biometric identification (such as facial recognition) in public spaces, and those used for social scoring (ranking individuals based on their personal characteristics, socio-economic status, or behaviour).

High-risk – AI applications that are expected to pose significant threats to health, safety, or the fundamental rights of persons. Notably, AI systems used in health, education, recruitment, critical infrastructure management, law enforcement or justice. They are subject to quality, transparency, human oversight and safety obligations, and in some cases require a Fundamental Rights Impact Assessment before deployment. A Fundamental Rights Impact Assessment (FRIA) is an ex ante review to identify and mitigate potential impacts on fundamental rights before an AI system is deployed. Earlier work on algorithmic impact assessments has suggested that such tools should identify which individuals and communities are affected by an automated system, describe possible harms, and provide a basis for public and institutional scrutiny of its use. They must be evaluated both before they are placed on the market and throughout their life cycle. The list of high-risk applications can be expanded over time, without the need to modify the AI Act itself. Citizens also have a right to submit complaints about AI systems and to receive explanations of decisions made by high-risk AI that affect their rights.

Limited risk – AI systems in this category have transparency obligations, ensuring users are informed that they are interacting with an AI system and allowing them to make informed choices. This category includes, for example, AI applications that make it possible to generate or manipulate images, sound, or videos (like deepfakes).

Minimal risk – This category includes, for example, AI systems used for video games or spam filters. Most AI applications are expected to fall into this category. These systems are not regulated, and Member States cannot impose additional regulations due to maximum harmonisation rules. Existing national laws regarding the design or use of such systems are overridden. However, a voluntary code of conduct is suggested.

The risk-based scheme follows a product-safety model in which regulatory duties are assigned to the providers and deployers of AI systems, and these duties become more demanding as the potential impact on health, safety, or fundamental rights increases. This structure is meant to ensure that oversight focuses on systems likely to create significant risks while allowing lighter approaches for uses considered less sensitive. Some legal scholars also argue that, in practice, the Act frames "trustworthy AI" as systems that can show compliance with these safety and risk thresholds. According to an initial appraisal by the European Parliamentary Research Service, the Commission's impact assessment drew on stakeholder consultations and a wide range of existing research when comparing policy options for this risk-based framework.

Added in 2023, the general-purpose AI category includes foundation models (for example, ChatGPT) that can perform a wide range of tasks. If a model's weights and design are made open source, developers must publish a training data summary and a copyright policy; closed-source models must meet broader transparency requirements. High-impact models that pose systemic risks (require more than 1025 floating-point operations to train) must undergo extra evaluation. A General-Purpose AI Code of Practice, published on 10 July 2025, outlines three main chapters on transparency, copyright, and safety and security to help providers demonstrate compliance with the AI Act. Participation in the code is voluntary.

Beyond these basic transparency duties, the Act sets a common list of obligations for providers of general-purpose AI models. They must publish a summary of the training data, adopt a policy to comply with copyright law, and provide technical documentation to downstream providers and supervisory authorities. Models that are designated as posing systemic risk must also carry out model evaluations and adversarial testing, assess and mitigate risks such as bias and security failures, report serious incidents, and ensure an adequate level of cybersecurity.

Exemptions…