Unacceptable Risk

AI systems deploying subliminal components beyond a person's consciousness or purposefully manipulative and deceptive techniques are prohibited when they materially distort behavior and cause or are likely to cause physical or psychological harm. This targets dark patterns amplified by AI that exploit cognitive biases and vulnerabilities, such as inaudible voice commands or imperceptible image flashes designed to influence purchasing decisions or political views without awareness.

Prohibits AI systems that exploit vulnerabilities of specific groups due to age or disability to materially distort behavior in a manner causing harm. This includes toys using AI to encourage dangerous behavior in children or voice assistants designed to manipulate elderly users with cognitive decline into financial transactions. The key legal test is whether the system's design targets the vulnerability to distort behavior, not merely whether harm occurs.

AI systems evaluating or classifying natural persons over time based on social behavior or personality characteristics are banned when the social score leads to:

• Detrimental treatment in unrelated social contexts
• Unjustified or disproportionate treatment This directly prohibits government-led social credit systems that rate citizens based on trustworthiness and restrict access to services like travel, education, or employment based on generalized behavioral patterns.

The use of real-time remote biometric identification systems in publicly accessible spaces for law enforcement purposes is prohibited with narrow exceptions. These systems match facial images against databases instantaneously without subjects' knowledge. Exceptions require prior judicial authorization and are limited to:

• Targeted search for specific crime victims (e.g., kidnapping)
• Prevention of imminent terrorist threat
• Localization of suspects for crimes punishable by at least 3 years imprisonment

AI systems making risk assessments of natural persons to predict the likelihood of committing a criminal offense are prohibited when based solely on profiling or personality traits. This bans pre-crime assessment tools that assign criminal risk scores to individuals based on demographic data, behavioral patterns, or psychological characteristics without objective, verifiable facts directly linked to criminal activity. Statistical correlations alone are insufficient legal grounds.

The creation or expansion of facial recognition databases through untargeted scraping of facial images from the internet or CCTV footage is prohibited. This directly bans companies like Clearview AI from harvesting billions of publicly available images to build biometric databases without consent. The prohibition addresses the chilling effect on privacy and freedom of assembly when individuals cannot appear in public without their biometric data being captured and indexed.

AI systems inferring emotions of natural persons are prohibited in workplace and educational settings, except for medical or safety purposes. This bans sentiment analysis cameras monitoring student engagement in classrooms or employee emotional states during performance reviews. The prohibition recognizes the scientific controversy around emotion detection validity and the power imbalance in employment and education contexts where refusal is not practically possible.

AI systems categorizing individuals based on biometric data to deduce or infer race, political opinions, trade union membership, religious beliefs, sex life, or sexual orientation are prohibited. This bans systems that claim to detect sexual orientation from facial features or political affiliation from gait analysis. The prohibition extends to both direct inference and proxy-based categorization where biometric data serves as the input for sensitive attribute classification.

The tier directly below prohibited practices. These systems are permitted but subject to mandatory conformity assessments before market placement. Examples include AI in critical infrastructure, educational assessment, employment screening, and law enforcement. Unlike unacceptable risk systems which are banned outright, high-risk systems must demonstrate compliance with harmonized standards for data governance, transparency, and human oversight.

A specific prohibited practice under Article 5(1)(c) of the EU AI Act. It refers to the evaluation or classification of natural persons over a period of time based on their social behavior or known personality characteristics, leading to detrimental treatment in unrelated contexts. The prohibition targets public authorities conducting mass surveillance-style scoring, not private credit scoring agencies operating under existing financial regulations.

The national enforcement body in each EU member state responsible for policing the AI Act's prohibitions. These authorities have the power to:

• Investigate suspected prohibited practices
• Order immediate withdrawal of non-compliant systems
• Impose fines up to €35 million or 7% of global annual turnover for unacceptable risk violations They coordinate through the European Artificial Intelligence Board to ensure consistent enforcement across the single market.

A mandatory pre-deployment analysis required for deployers of high-risk systems, but the concept originates from the need to identify when a system crosses into unacceptable risk territory. The assessment must document:

• The specific groups of individuals likely to be affected
• The risks to fundamental rights (dignity, non-discrimination, privacy)
• The duration and frequency of the system's use Systems found to pose a clear threat to fundamental rights during this assessment must not be deployed.

A mandatory obligation requiring providers to immediately notify market surveillance authorities of any malfunction or failure that leads to death or serious harm. While this applies to high-risk systems, the reporting mechanism serves as a safeguard against de facto unacceptable risk—if a system consistently causes severe harm, authorities can reclassify it and order its removal from the market. Reports must be filed within 72 hours of the provider becoming aware of the incident.