EU AI Act establishes risk-based rules and bans for AI systems

A risk-based regulatory framework

The EU AI Act organizes artificial intelligence into four distinct categories based on the potential impact of the technology. These categories are defined as unacceptable, high, limited, and minimal risk. The classification of a system depends entirely on its intended purpose rather than the underlying technology itself.

For example, a general software product is not automatically treated as an AI system under these rules. However, machine-learning software used for specific tasks like recruitment or credit assessment may fall into the high-risk category. This means the same piece of code could be subject to different rules depending on whether it is used for a simple task or a consequential human decision.

The Act applies to any organization that places an AI system on the EU market or puts it into service within the Union. This reach extends to providers outside of Europe if the output of their system is used within the EU. The framework is designed to ensure that safety and transparency are built into the entire lifecycle of the software.

Banned practices and protections

The regulation explicitly prohibits AI practices that pose unacceptable risks to human safety, rights, or the ability to make free choices. This includes systems that use subliminal techniques or deceptive methods to materially distort behavior. It also targets the exploitation of vulnerabilities linked to a person's age, disability, or specific social and economic situations.

Social scoring is another practice that is strictly banned if it leads to unjustified or disproportionate detrimental treatment. The Act also restricts certain biometric applications, such as the untargeted scraping of facial images to build recognition databases. Additionally, it prohibits biometric categorization that infers sensitive traits like political opinions or sexual orientation.

In specific settings like workplaces and schools, the Act restricts the use of AI to infer emotions from biometric data. This restriction does not apply to ordinary identity verification or systems used for medical or safety reasons. These bans are intended to prevent the use of AI in ways that could lead to significant harm or the erosion of individual privacy.

Requirements for high-risk systems

Systems classified as high-risk, such as those used in education, employment, and critical infrastructure, face rigorous obligations. Providers must implement a continuous risk-management system to identify and mitigate foreseeable risks. They are also required to ensure high-quality training, validation, and testing datasets that are free from bias.

Deployers, who are the organizations using these tools in a professional setting, must also follow specific rules. They are responsible for assigning human oversight to competent people who can interpret outputs and intervene if necessary. If a deployer makes substantial modifications to a tool, the responsibilities of the original provider may shift to the deployer.

Before a high-risk system can be sold, it must undergo a conformity assessment and receive a CE marking. Providers must also maintain a post-market monitoring plan to track performance and report serious incidents. This ensures that authorities can inspect the system's technical documentation and trace how specific outputs were produced from input data.

General-purpose AI and foundation models

The Act also addresses general-purpose AI models, which are systems capable of performing a wide range of distinct tasks. This includes large language models and image generators, whether they are provided through an API or as downloadable weights. Providers of these models must maintain technical documentation and provide a summary of the content used for training.

Copyright compliance is a key requirement for these models before they are placed on the EU market. Providers must adopt policies to respect rights reservations and honor machine-readable opt-outs during data collection. While some documentation requirements are exempt for open-source models, the training-content summary remains a mandatory requirement for all.

Models that present systemic risks, such as those with high-impact capabilities, face even stricter duties. These include conducting adversarial testing and reporting serious incidents to the AI Office. These extra safeguards are designed to ensure that the most powerful models do not pose a threat to the security or rights of the general public.

Enforcement and next steps

The EU AI Act is being implemented in a phased approach rather than all at once. Prohibitions and AI-literacy duties began in February 2025, while governance rules for general-purpose AI followed in August 2025. Most remaining provisions, including those for many high-risk systems, will apply starting in August 2026.

Enforcement is handled by national authorities and EU-level bodies, such as the AI Office. These bodies have the power to conduct inspections, request technical documentation, and order corrective actions. They can also restrict a system's availability or require it to be withdrawn from the market entirely if it violates the law.

Penalties for non-compliance are significant, particularly for prohibited practices. Violations can result in administrative fines of up to €35 million or 7% of a company's worldwide annual turnover. Businesses are encouraged to inventory their current AI tools and identify which risk categories they fall into to ensure they meet these upcoming requirements.