EU AI Act

AI practices deemed a clear threat to fundamental rights are strictly prohibited. This tier bans systems that deploy subliminal manipulation, exploit vulnerabilities of children or persons with disabilities, or implement social scoring by public authorities. Real-time remote biometric identification in publicly accessible spaces by law enforcement is also banned, with narrow exceptions for serious crime.

Systems posing significant potential harm to health, safety, or fundamental rights. This tier includes AI used in critical infrastructure, educational scoring, employment management, essential services eligibility, law enforcement, migration, and democratic processes. Providers must implement a risk management system, compile technical documentation, undergo conformity assessment, and maintain post-market monitoring.

AI systems subject to transparency obligations to ensure informed human interaction. This tier covers chatbots, emotion recognition systems, and deepfake generators. Users must be clearly informed they are interacting with an AI system unless it is obvious. Providers of AI-generated synthetic content must label outputs as artificially generated or manipulated.

The vast majority of AI systems currently in use fall into this category, including AI-enabled video games and spam filters. These systems face no mandatory regulatory obligations under the AI Act. Providers may voluntarily adopt codes of conduct to demonstrate best practices, but no conformity assessment or documentation is required for market placement.

A mandatory, iterative process that must be planned and documented throughout the entire AI system lifecycle. Providers must identify reasonably foreseeable risks to health, safety, and fundamental rights, estimate their severity and probability, and implement appropriate mitigation measures. Residual risks must be deemed acceptable before the system can be placed on the market.

• Must be continuously updated based on post-market monitoring data
• Includes analysis of risks arising from reasonably foreseeable misuse
• Requires testing to identify the most effective risk mitigation measures

Training, validation, and testing datasets are subject to rigorous governance requirements. Providers must examine datasets for potential biases that could lead to discrimination, ensure data is relevant and representative of the intended population, and assess for errors and completeness. The data must be appropriate for the system's intended purpose and geographical context.

• Mandatory examination for statistical bias and prohibited discriminatory patterns
• Data provenance and lineage must be documented
• Special rules apply when processing special categories of personal data to mitigate bias

Before placing a system on the market, a provider must compile a comprehensive dossier demonstrating compliance. This includes a detailed description of the system's design, development, and performance characteristics, including its architecture, algorithmic logic, and the metrics used to measure accuracy and robustness. The documentation must be kept for 10 years and be readily accessible to authorities.

• Must include instructions for use and information on human oversight mechanisms
• Describes the system's intended purpose and foreseeable unintended consequences
• Forms the basis for the conformity assessment by a notified body

High-risk systems must be designed to allow for effective human oversight, preventing automation bias and ensuring decisions are not blindly followed. The human operator must have the competence, authority, and actual capacity to understand the system's outputs, override automated decisions, and intervene in real-time. This goes beyond a simple 'rubber stamp' on an AI recommendation.

• Operators must be able to interpret and question the system's output
• Systems must include built-in controls for meaningful human intervention
• Oversight logs must be maintained to audit operator actions and situational awareness

Compliance does not end at deployment. Providers must establish a continuous, systematic process to collect and analyze real-world performance data. This system must proactively identify emerging risks, unexpected interactions, and drift in model accuracy. Findings must feed back into the risk management system and can trigger mandatory serious incident reporting to market surveillance authorities.

• Requires a documented post-market monitoring plan
• Must capture data on system performance across diverse demographic groups
• Triggers immediate corrective action if a previously unidentified risk emerges

Providers must implement a formalized, documented organizational structure of policies, processes, and procedures to ensure consistent compliance. This includes a clear strategy for regulatory adherence, robust design control techniques, and procedures for managing substantial modifications. The QMS ensures that compliance is an institutional capability, not a one-time project.

• Covers the entire product lifecycle from design to decommissioning
• Must include procedures for vendor and sub-contractor management
• Subject to audit by notified bodies during the conformity assessment

The Act distinguishes between two primary actors with distinct legal duties:

Providers (developers placing AI on the market):

• Implement a risk management system throughout the lifecycle
• Compile exhaustive technical documentation
• Conduct conformity assessments and affix CE marking
• Establish post-market monitoring and report serious incidents

Deployers (professional users):

• Ensure human oversight with meaningful intervention capability
• Conduct fundamental rights impact assessments
• Monitor operation for signs of malfunction or drift
• Retain automatically generated logs for audit purposes

A dedicated regime for General Purpose AI (GPAI) models trained on broad data at scale:

• All GPAI models must provide technical documentation and comply with EU copyright law by publishing a summary of training data.
• Models classified as posing systemic risk—determined by cumulative compute exceeding 10^25 FLOPs or Commission designation—face additional obligations:
  • Mandatory adversarial testing and model evaluation
  • Serious incident reporting to the AI Office
  • Cybersecurity protections and energy consumption disclosure
• The AI Office serves as the central enforcement body for GPAI regulation.

High-risk AI systems must demonstrate compliance before market entry through one of two routes:

• Internal Control: The provider self-assesses conformity when applying harmonized standards that provide a presumption of conformity. This is the default pathway for most high-risk systems.
• Notified Body Assessment: Required for specific high-risk categories (biometric systems, safety components) where an accredited third-party notified body audits the quality management system and technical documentation.

A substantial modification to an already-certified system triggers a new conformity assessment cycle.

The Act creates a robust enforcement architecture with significant financial consequences:

• Market Surveillance Authorities in each member state investigate non-compliance and can demand corrective action or restrict market access.
• Penalty structure based on infringement severity:
  • Up to €35 million or 7% of global annual turnover for prohibited practices
  • Up to €15 million or 3% for most other violations
  • Up to €7.5 million or 1.5% for supplying incorrect information
• Regulatory sandboxes allow supervised testing of innovative systems without immediate penalty exposure.
• The EU AI Board coordinates enforcement across member states to ensure consistent application.

A core pillar of the Act is the obligation to provide clear, accessible information:

• Technical documentation must detail system architecture, training methodologies, data provenance, and performance metrics—maintained for 10 years.
• Automated profiling and consequential decisions require explicit notification to affected individuals with a right to meaningful human intervention.
• Human oversight logs must capture override events, operator situational awareness, and intervention rationale.
• Emotion recognition and biometric categorization systems must inform exposed individuals in real-time.
• Deepfake content must be labeled as artificially generated or manipulated.