European Union - Prohibited AI Practices Guidelines

Overview

The European Commission's "Guidelines on prohibited artificial intelligence (AI) practices" (published 4 February 2025) interpret and operationalise Article 5 of Regulation (EU) 2024/1689 (the AI Act). The Guidance identifies and explains the cumulative conditions that make an AI application fall within the AI Act's prohibited category (for example, manipulation using subliminal techniques, exploitation of vulnerabilities, social scoring, biometric scraping and emotion inference in workplaces or educational settings). The document aims to promote consistent application across Member States by providers, deployers and national competent authorities. The Guidelines are non‑binding; the authoritative legal interpretation rests with the Court of Justice of the European Union, but the Commission expects national authorities to rely on these clarifications. For the official publication and downloads see the Commission library page: Commission: Guidelines on prohibited AI practices (4 Feb 2025) and the Commission redirection to the English guideline document: Guidelines on prohibited artificial intelligence practices (English).

Definitions

The Guidelines provide working definitions to reduce ambiguity: "materially distorting behaviour" is framed as influence that appreciably impairs a person's ability to make an informed, autonomous decision and causes or is reasonably likely to cause "significant harm" (which includes physical, psychological, financial, social or legal harms). "Subliminal techniques" are defined as inputs beyond a person's conscious perception; "purposefully manipulative or deceptive techniques" cover engineered persuasion tactics that go beyond lawful persuasion and lead to appreciable impairments. "Vulnerabilities" are understood functionally (e.g., cognitive, developmental, socio-economic) and not limited to medical conditions. "Real-time" denotes negligible delay between capture and identification; "publicly accessible spaces" include streets and open squares but exclude controlled-access environments such as prisons and certain online spaces. These working definitions guide the cumulative tests applied to each Article 5 subparagraph.

Governance and Institutional Framework

The Guidelines situate Article 5 within the AI Act's phased application: Chapters I and II (including Article 5) have applied since 2 February 2025, while governance, enforcement and penalties fully applied as of 2 August 2025. Member States must designate national market surveillance authorities and notify the Commission; the Commission's European AI Office and the European AI Board coordinate implementation, provide further guidance and support consistency across Member States. Enforcement actors include national competent authorities, market surveillance authorities, data protection authorities (where GDPR issues intersect), the European Data Protection Supervisor (for Union institutions) and courts. The Guidance stresses cooperation and information exchange, and references the AI Act's requirement that Member States notify national rules on penalties and enforcement measures to the Commission. See the AI Act authoritative text at EUR-Lex: Regulation (EU) 2024/1689 (AI Act) — EUR-Lex and the AI Act Service Desk explanatory pages at AI Act Service Desk – Article 5.

Key Focus Areas

The Guidelines break Article 5 into discrete prohibited practices and explain each with cumulative conditions and examples. Key focus areas include: (1) Manipulation and deception (Art. 5(1)(a)) — AI that uses subliminal or engineered persuasive techniques to materially distort behaviour; (2) Exploitation of vulnerabilities (Art. 5(1)(b)) — targeting persons because of age, disability or socio‑economic status; (3) Social scoring (Art. 5(1)(c)) — assigning social reliability scores leading to unjustified discrimination; (4) Predictive criminal risk assessments (Art. 5(1)(d)) — profiling individuals to predict criminality solely on automated inference; (5) Untargeted scraping for facial recognition databases (Art. 5(1)(e)) — mass collection of biometric images from the internet or CCTV without consent or a lawful basis; (6) Biometric categorisation of protected attributes (Art. 5(1)(g)) — systems inferring race, religion, political opinions, sexual orientation, trade union membership and similar sensitive traits; (7) Emotion inference in workplaces and educational institutions (Art. 5(1)(f)) — inference of a specific individual's emotions (distinct from aggregate mood analysis), except narrowly for medical or safety reasons; and (8) Real-time remote biometric identification in public spaces for law enforcement (Art. 5(1)(h)) — banned except for strict, narrowly defined exceptions like locating missing persons or preventing an imminent threat subject to authorisation and oversight. The Guidelines detail the interplay of objective and effect-based tests and provide use case examples to help providers assess compliance.

Implementation Framework

The Guidance recommends an implementation approach for providers and deployers: (1) early legal and technical screening to determine whether an AI system falls within Article 5 prohibitions; (2) lifecycle documentation and records to show decisions and assessments; (3) obligations to cease placing on the market, to withdraw, or to prevent use where a system constitutes a prohibited practice; (4) cooperation with national competent authorities, including providing requested information and, where required, assistance in mitigation or withdrawal; and (5) embedding AI literacy and governance measures among staff and contractors. The document emphasises that even where a practice is not expressly listed in Article 5, providers should evaluate whether a given system meets the cumulative criteria (objective/effect, material distortion, significant harm) and should take a precautionary approach where uncertainty exists. The guidelines also identify permitted narrow exceptions (medical/safety, strictly necessary law enforcement cases with authorisation, research/testing under safeguards, and certain tagging or annotation activities) and clarify that free/open-source development may still be captured if actual deployment constitutes a prohibited practice.

Monitoring and Evaluation

The Guidelines set out monitoring expectations: Member States' market surveillance authorities and national competent supervisory bodies are to perform post-market monitoring and cooperate through the European AI Board and the AI Office. The Commission recommends periodic reviews of emerging use cases and invites national authorities to report trends, enforcement actions, and open issues to inform updates. Providers and deployers are encouraged to maintain post-market monitoring logs, incident reporting mechanisms and to perform regular audits for inadvertent drift that could transform a permitted system into one producing prohibited effects. The Guidance also asks national bodies to collect evidence-based case studies to refine legal interpretation and to feed into potential future delegated or implementing acts where needed.

Penalties, Liability, and Appeals

The AI Act's penalties framework is summarised and linked to the Guidelines: non-compliance with Article 5 is subject to the top-tier administrative fines (up to EUR 35,000,000 or up to 7% of the offender's total worldwide annual turnover, whichever is higher). Member States must implement effective, proportionate and dissuasive penalties, and formally notify the Commission of national penalty rules. The Guidelines explain that in addition to fines, authorities may order corrective measures (withdrawal, suspension of operations, orders to bring systems into conformity) and criminal or civil liability may follow under national law. The Guidance details defence and procedural safeguards, including rights to be heard and appeal routes before national courts and, where applicable, referral to the CJEU for legal questions. It notes the GDPR interplay where biometric or sensitive personal data processing also triggers supervisory authority actions and penalties under data protection rules.

Relationship to Other Instruments

The Commission explains how Article 5 interacts with existing EU instruments. The Guidelines stress compliance with the GDPR (Regulation (EU) 2016/679) for personal data processing and with the Law Enforcement Directive where processing is for law‑enforcement purposes; they also reference Union non‑discrimination law, sectoral safety rules (e.g., medical devices, transport safety), and product/supply chain safety legislation. The Guidance clarifies that where other legal regimes already prohibit or regulate certain practices (e.g., national bans on some biometric uses), Article 5 operates alongside and may reinforce those protections. It also underscores that national security, defence, certain judicial cooperation uses and some research activities are excluded from the AI Act's scope, per the Regulation's exclusions.

International Alignment

The Guidelines encourage international cooperation and note that the EU intends these interpretations to inform global approaches. The Commission invites exchanges with partners, standardisation bodies and multilateral fora to promote convergence on unacceptable AI practices and to avoid regulatory fragmentation that could undermine human rights protections. The document references the need to align, where possible, with international human‑rights standards and comparable regulatory developments (including OECD AI principles, Council of Europe work, and multilateral dialogues). However, the Guidance clarifies that extraterritorial application of Article 5 means third-country providers placing AI systems on the EU market or making them available in the Union must comply with the prohibitions.

Implementation Timeline

Sources and References