EU AI Act - Prohibited Practices Phase 1

Applied February 2025. Bans unacceptable-risk AI: social scoring, biometric identification in public spaces, manipulation, and exploitation of vulnerabilities. Fines up to 35 million EUR.

Jurisdiction	European Union
Category	AI Regulations
Status	Active
Effective date	2025-02-02
Latest development	2026-07-01
Enforcement actions on record	1

Analysis

The EU AI Act’s prohibited practices (Article 5 / “unacceptable-risk AI”) have been fully applicable since 2 February 2025, with violations subject to administrative fines up to EUR 35 million or 7% of worldwide annual turnover and immediate withdrawal orders for non-compliant AI systems.3 8

Key Requirements (EU AI Act – Prohibited Practices, Phase 1)

Under Article 5 of Regulation (EU) 2024/1689 (EU AI Act), certain AI practices are categorically banned as “unacceptable risk,” meaning they cannot be placed on the market, put into service, or used in the EU.3 6 7

Key prohibited practices (effective 2 February 2025) include:3 5 6 8

Harmful AI-based manipulation and deception
AI systems deploying subliminal, purposefully manipulative, or deceptive techniques beyond a person’s consciousness that materially distort behaviour and are likely to cause significant harm.1 3 6
Detailed conditions and examples are set out in the Commission Guidelines on Prohibited AI Practices.1 2

Exploitation of vulnerabilities
AI systems that exploit vulnerabilities of specific groups (e.g., age, physical or mental disability, or specific social or economic situation) with the objective or effect of materially distorting behaviour and causing or likely to cause significant harm.3 5 6
This includes targeting children, the elderly, or persons with disabilities in ways that leverage their vulnerability.1 6

Social scoring by public authorities
AI systems that evaluate or classify natural persons or groups based on social behaviour, known or predicted personal or personality traits, or socio-economic status, leading to detrimental or disproportionate treatment that is unjustified or unrelated to the context.3 5 6
The ban specifically targets social scoring by public authorities that affects rights and access to services.5 8

Real-time remote biometric identification in publicly accessible spaces for law enforcement (with narrow exceptions)
Real-time remote biometric identification systems (typically facial recognition) used by law enforcement in publicly accessible spaces are prohibited, except for narrowly defined use cases (e.g., searching for specific crime victims, preventing imminent terrorist attacks, or identifying perpetrators of serious crimes).3 5 8
Conditions and exceptions are detailed in Article 5 and related recitals of the AI Act text.7

Biometric categorisation inferring protected characteristics
Biometric categorisation systems that use biometric data (e.g., facial images, gait) to infer or classify individuals by protected characteristics such as race, political opinions, trade union membership, religious or philosophical beliefs, sex life or sexual orientation are banned.3 5 6

Emotion recognition in workplaces and educational institutions
AI systems that infer emotions of natural persons in employment or educational settings (e.g., monitoring workers’ emotional states or students’ emotions) are prohibited.3 5 8

Untargeted scraping of images/video for facial recognition databases
Untargeted scraping of images or video from the internet or CCTV feeds to create or expand facial recognition databases is banned.3 5 8

Individual criminal offence risk assessment or prediction
AI systems that assess or predict the risk of natural persons committing criminal offences, based primarily on profiling or past criminal behaviour, are prohibited due to risks of discrimination and fundamental rights violations.3 5 8

AI systems generating NCII/CSAM (non-consensual intimate images / child sexual abuse material)
AI systems used to generate or manipulate image, audio, or video content that constitutes non-consensual intimate images (NCII) or child sexual abuse material (CSAM) are expressly prohibited.8 3

Scope and enforcement
These prohibitions apply to providers and deployers of AI systems in the EU, regardless of where they are established.3 6
National market surveillance authorities and the EU AI Office can investigate, order withdrawal, and impose fines.3 8
Full text: EUR-Lex – Regulation (EU) 2024/1689 (AI Act)7

Compliance Challenges

Organizations face several recurring compliance challenges when addressing Article 5 prohibitions:2 3 4

Identifying whether a system falls under a prohibited category
Many systems (e.g., recommendation engines, engagement optimizers, behavioural analytics) may border on manipulative or exploitative practices, requiring nuanced legal and technical assessment.1 4
The Commission’s Guidelines on Prohibited AI Practices were published specifically to address interpretive challenges for Article 5.1 2

Legacy and global systems
Multinational companies often operate global AI platforms whose functionality in the EU must be altered (e.g., disabling emotion recognition for EU users while maintaining global features).3 5
EY notes organizations struggle to map existing AI systems to the AI Act’s risk categories and prohibited practices.4

Data practices and biometric scraping
Companies that have built facial recognition datasets from broad web scraping must reassess legality and provenance, potentially retiring entire datasets used for training.3 8
Industry analyses highlight major impact on vendors offering large-scale biometric identification solutions based on scraped data.8

Law enforcement and public sector adaptation
Police and public authorities need to transition away from real-time facial recognition in public spaces, which has been piloted or deployed in several EU jurisdictions, now constrained to narrow exceptions under the AI Act.3 6
The Commission guidance explains operational limitations and lawful exception scenarios.1 2

Organizational governance and awareness
Many organizations lack AI-specific governance structures and internal controls to systematically detect and prevent prohibited AI uses.3 4
Universities and research institutions, such as Harvard’s EU AI Act guidance, emphasize challenges in mapping academic and experimental AI systems to Article 5 constraints.5

Real examples and case/industry reports

The EY “AI Act: Prohibited AI Practices Become Applicable” alert gives examples of:
AI-based subliminal manipulation in marketing and behavioural nudging tools.
Predictive policing systems that fall under the prohibition on criminal offence risk prediction.4
The Harvard University EU AI Act guidance lists specific prohibited use cases in employment and education, illustrating real organizational scenarios (e.g., emotion recognition in HR or student monitoring systems).5
The Commission’s Guidelines provide hypothetical and sector-specific examples (e.g., manipulative “dark patterns” powered by AI, child-targeted recommendation engines) to clarify what constitutes prohibited exploitation or manipulation.1 2

Implementation Best Practices

To comply with the Prohibited Practices Phase 1, organizations should establish structured governance and technical controls aligned with the AI Act’s risk-based framework.3 4 1

Actionable steps

Comprehensive AI inventory and classification
Map all AI systems (including models, tools, and embedded functionalities) and classify them by AI Act risk category, explicitly checking for Article 5 prohibited patterns.3 4
Use the Commission’s Guidelines on Prohibited AI Practices as a checklist to detect manipulative, exploitative, biometric, and social scoring features.1 2

Policy and design constraints
Develop internal AI design policies that categorically forbid:
Social scoring functionality for public authority clients.
Emotion recognition modules in HR/education deployments.
Untargeted scraping for biometric datasets.
Predictive policing or criminal risk scoring.3 5 8
Incorporate “prohibited patterns” checks into product development life cycles and model governance workflows.3 4

Data governance and biometric safeguards
Audit training and operational datasets for untargeted biometric scraping or data used for NCII/CSAM-related purposes, removing unlawful data sources.3 8
Align data practices with GDPR (lawfulness, purpose limitation, special categories of data) as part of AI compliance.9[3](https://digital-strategy.ec.europa.eu/en/policies/regulatory

Recent developments

2026-07-01 — Regulatory and compliance resources continue to refine guidance on **Article 5 interpretation**, clarifying that both “subliminal” and “purposefully manipulative or deceptive” AI techniques that materially distort decision‑making and cause or are likely to cause significant harm are outright prohibited, influencing product‑design and risk‑assessment practices across the AI industry.[5] (source)
2026-06-18 — University and research IT governance teams highlight the **operational implications** of the prohibited‑use list, advising institutions to avoid deploying campus systems that perform emotion inference, real‑time biometric identification in public spaces, or social‑scoring‑like assessments, to remain compliant with Phase 1 of the EU AI Act.[4] (source)
2026-06-10 — Specialist EU‑AI‑Act commentators note growing **market impact**: vendors are removing emotion‑recognition modules from HR and EdTech products, discontinuing untargeted facial‑image scraping, and revising biometric categorisation functionalities, as these uses fall squarely within the banned Article 5 practices.[3] (source)
2026-06-05 — Industry legal analysis underscores that the initial prohibitions apply **extraterritorially** to providers and deployers worldwide whenever AI outputs are used in the EU, prompting global AI companies to disable or redesign features such as social scoring, subliminal manipulation, and certain predictive‑policing tools to avoid EU‑level sanctions.[2] (source)
2026-06-01 — Privacy and compliance experts report that Article 5 **prohibited practices are now actively enforced**, highlighting regulatory focus on manipulative techniques, social scoring, criminal‑prediction software, emotion recognition in workplaces/education, and exploitation of vulnerable groups, with potential fines up to €35 million or 7% of global turnover.[6][2] (source)
2026-05-20 — Commission guidance on the phased application of the AI Act confirms prohibited practices and AI‑literacy obligations as the **first wave of binding rules**, preceding governance and GPAI model obligations (from 2025‑08‑02) and high‑risk AI requirements, thereby framing “Prohibited Practices Phase 1” as an immediate compliance priority for any AI deployed in the EU.[1] (source)
2026-05-15 — The European Commission reiterated that the **eight prohibited AI practices** (harmful manipulation, exploitation of vulnerabilities, social scoring, predictive criminal‑risk assessment, untargeted facial‑image scraping, emotion recognition at work and school, biometric categorisation for protected traits, and real‑time remote biometric ID for law enforcement) have been applicable since 2025‑02‑02, emphasizing enforcement as the AI Act’s first operational phase.[1] (source)
2026-05-10 — The Omnibus package confirms that existing prohibited practices will be **supplemented from 2026‑12‑02** by rules targeting AI‑generated intimate content and CSAM, signaling that providers of generative image/video tools must treat certain sexual‑content use cases as categorically banned AI under the Act.[7] (source)
2026-05-08 — Under the Omnibus agreement, the **scope and timelines** of the EU AI Act are adjusted: deadlines for certain high‑risk AI obligations and watermarking requirements are extended, while the core Article 5 prohibitions on unacceptable‑risk AI (manipulation, exploitation of vulnerabilities, social scoring, etc.) remain already in force, reinforcing Phase 1 compliance expectations.[7] (source)
2026-05-07 — EU institutions reached political agreement on the “AI Act Omnibus,” which, among other changes, introduces a **new prohibition** on AI “nudifier” applications that generate or manipulate sexually explicit or intimate content, including CSAM, extending the prohibited practices regime beyond the original eight categories.[7] (source)

Enforcement actions

Date	Authority	Entity	Fine	Summary
2025-11-03	European AI Office	Clearview AI	30,500,000 EUR	Prohibited real-time biometric identification system used by law enforcement clients in EU. (source)

Related regulations

Internet Information Service Algorithmic Recommendation Management Provisions — China, Active
National Artificial Intelligence Policy Bill — Chile, Active
Artificial Intelligence Liability Directive — European Union, Active
Federal Law on the Development and Use of Artificial Intelligence — Mexico, Active
Presidential Regulation on the Framework for Artificial Intelligence Development — Indonesia, Active
Law Establishing the Framework for the Development and Use of Artificial Intelligence — Peru, Active
California Consumer Privacy Act AI Amendments — California, Active, effective 2020-01-01
GPAI Framework for Responsible AI — AI Regulations, Active

Put it into practice

Generate the policy: EU AI Act policy generator (generatepolicy.com)
Buy the policy pack: EU AI Act Compliance Policy (cyberpolicy.shop)
Build it yourself: Pillar 06 Companion — The 2026 AI Risk Register (ciso.diy)

Browse all regulations · Compliance deadlines · Latest updates