EU sets out guidance on banning harmful AI uses

Mass surveillance, detecting emotions, social scoring: EU regulators have laid out what types of artificial intelligence tools are to be outlawed as too dangerous under the bloc’s pioneering AI Act.Provisions banning uses of the technology deemed to pose an unacceptable risk came into force this week, although the European Union’s 27 states will have until August to designate a regulator to enforce them.Adopted last year, the EU rules aim to avoid abuses of the nascent technology without hobbling innovation in Europe, which faces an uphill challenge as the United States and China charge ahead in the AI field.The law, considered the most comprehensive AI regulation in the world, takes a risk-based approach: if a system is high-risk, a company will have a stricter set of obligations to fulfil before being authorised in the EU.The new guidance aims to provide legal clarity to both regulators and companies on what uses face an outright ban, said a senior official with the European Commission.The official stressed that the non-binding guidance aims to offer "insights” into the commission’s understanding of the bans, and that the Court of Justice of the European Union will set authoritative interpretations of the new law.Companies in violation of the rules face fines of up to seven percent of worldwide annual revenue, or fines of up to 35 million euros ($37 million) – whichever is higher.The commission set out eight cases justifying a ban:1. Identifying people in real time using cameras in public spacesThe law bans using camera-equipped AI systems for real-time biometric identification in public spaces for law enforcement purposes. The aim is to prevent scenarios where an individual could be scooped up by police without cross-checking with real-world information. Exceptions exist, such as preventing specific threats including terrorist attacks.2. Social scoring based on personal data unrelated to riskThe rules bar using AI to rank people’s likelihood of defaulting on a loan or committing social welfare fraud, for instance, by relying on personal data unrelated to the context – such as origin, skin colour, or behaviour on social networks.3. Evaluating criminal risk based on biometric data alonePolice cannot use AI to predict an individual’s risk of criminal behaviour, such as participating in riots or committing an attack, based solely on facial features or other personal characteristics, without taking into account objective and verifiable facts directly related to their actions.4. Scraping online images to create facial recognition databasesThe AI Act bans tools that scour the Internet and CCTV footage to indiscriminately extract facial images and create large-scale databases of billions of images — which would amount to state surveillance.5. Emotion detection in workplaces and places of educationOrganisations are barred from using webcams or voice recognition systems to detect people’s emotions.6. Behaviour manipulation using AIIntegrating deceptive or subliminal AI systems into the design of an interface to push users to make a purchase is banned.7. Exploiting vulnerabilities related to age or disabilityThe rules ban AI-based toys or other systems aimed at children, the elderly or otherwise vulnerable people, that are designed to encourage potentially-harmful behaviour.8. Inferring political opinions or sexual orientation based on biometric dataSystems that claim to infer people’s political opinions or sexual orientation from analysing their face will not be allowed in the EU under the AI Act.