Which AI practices are prohibited under the EU AI Act?

Updated 9 August 2026 for Regulation (EU) 2026/1744.

Article 5 now bans ten categories of AI practice outright. Eight have applied since 2 February 2025: harmful manipulation, exploiting vulnerabilities, social scoring, predictive policing on profiling alone, untargeted facial-image scraping, workplace and school emotion recognition, sensitive biometric categorisation, and real-time remote biometric identification. Two more — non-consensual intimate imagery and AI-generated child sexual abuse material and performances — are new points (ba) and (bb), applying from 2 December 2026 (Article 113, third paragraph, point (a), as amended by Regulation (EU) 2026/1744).

Quick answer

• Two dates, one point: Chapters I and II — including the Article 5 prohibitions — apply from 2 February 2025 under Article 113, third paragraph, point (a). That point now excepts Article 5(1), first subparagraph, points (ba) and (bb), and Article 5(1a) and (1b): those apply from 2 December 2026.
• What's banned: Ten practices — the original eight, from manipulative and exploitative systems to social scoring, scraping, workplace emotion recognition and biometric categorisation, plus intimate imagery and child sexual abuse material and performances.
• Who it binds: most points cover placing on the market, putting into service, and use — though (d) to (g) qualify the middle limb as "the putting into service for this specific purpose", and (h) reaches use alone. Unlike most of the Act, Article 5 binds providers and deployers alike — including deployers who merely use a banned system.
• The penalty: Up to €35 million or, "if the offender is an undertaking", 7% of total worldwide annual turnover for the preceding financial year, whichever is higher — the Act's highest tier (Article 99(3)). For SMEs, including start-ups, Article 99(6) inverts it to "whichever thereof is lower".
• Enforcement caveat: Chapter XII, which carries the penalties, only applied from 2 August 2025 (Article 113, third paragraph, point (b)) — so national enforcement machinery lagged the prohibition itself by six months.

What exactly does Article 5 prohibit?

Ten practices. Eight have been prohibited since 2 February 2025; Regulation (EU) 2026/1744 inserted points (ba) and (bb), which apply from 2 December 2026 and which Articles 5(1a) and (1b) narrow before they reach anyone. The list itself — every point in the Regulation's own terms, with the carve-outs each one carries — is the ten prohibited practices, point by point. This page is about which of them reach an ordinary software company, and what to do about it.

Two things travel with the list wherever it is read. Article 5(8) says the prohibitions do not displace other bans that apply where an AI practice infringes other Union law — GDPR, consumer-protection and non-discrimination law still bite independently. And the ban is on a practice, not on a tier: it reaches placing on the market, putting into service and use, so a deployer who buys a prohibited system has not moved the problem onto the vendor.

The first version of our own classifier could not return "prohibited" at all — every banned practice landed in a reassuring tier, workplace emotion recognition included, until we split Article 5 into its fact-based prohibitions, which a questionnaire can detect, and its judgement-based ones, which it must flag for legal review instead of deciding. If your classification process has no path to "this practice must stop", it is not classifying — it is reassuring.

Which of these actually apply to a B2B software company?

Most bans target law-enforcement and public-authority use. Six are the realistic exposure for SaaS, AI, fintech, healthcare and HR-tech vendors selling into the EU:

• (a) manipulation — relevant if your product uses persuasive or "dark pattern" design that could be argued to distort decisions and cause significant harm.
• (b) exploiting vulnerabilities — relevant where a system targets users by age, disability, or economic distress (e.g. debt or subprime-lending flows).
• (ba) and (bb) generated intimate imagery, CSAM and performances — from 2 December 2026, and not sector-specific: the exposure is real for anyone shipping image, video or audio generation. Intent is not the only trigger — Article 5(1a), point (a)(ii) also reaches inadequately safeguarded systems.
• (f) emotion recognition — the sharpest one for HR tech: AI that infers a candidate's or employee's emotional state from voice or video in a hiring, monitoring or performance context is prohibited, not merely high-risk. The narrow carve-out is medical or safety use.
• (g) biometric categorisation — relevant to any analytics that infer the sensitive characteristics listed above from biometric data, but not to labelling or filtering a lawfully acquired biometric dataset, which the point expressly excludes.

Predictive policing (d), untargeted scraping (e) and RBI (h) mainly concern law enforcement, but (e) also catches any commercial actor building a face database by scraping.

The Commission published guidelines on prohibited practices in early 2025 to help interpret these boundaries. Treat them as guidance, not law — they clarify the Commission's reading of Article 5 but do not change the binding text, and the Court of Justice has the final word on interpretation.

How does 2 February 2025 relate to the rest of the timeline?

It was the first of the staggered application dates. Entry into force was 1 August 2024, but no substantive obligation applied then. Article 113, third paragraph, point (a) switched on Chapters I and II — the general provisions, definitions, the AI-literacy duty (Article 4), and the Article 5 prohibitions — from 2 February 2025. Point (b) followed on 2 August 2025, covering Chapter III Section 4, Chapter V, Chapter VII and Chapter XII and Article 78 — with the exception of Article 101. The Regulation as a whole applies from 2 August 2026.

Two dates now sit ahead of you. The new prohibitions bite on 2 December 2026. And the high-risk requirements and operator obligations in Chapter III, Sections 1, 2 and 3 no longer apply from 2 August 2026: they apply from 2 December 2027 for Annex III systems and 2 August 2028 for Annex I systems (Article 113, third paragraph, point (c), as amended by Regulation (EU) 2026/1744). Every date: the EU AI Act timeline.

What this means for you

If you're a provider: Article 5 is not a documentation exercise — there is no conformity assessment that makes a prohibited practice compliant. You must remove the capability. Run a focused audit of any feature touching emotion inference, biometric categorisation, persuasive design, or vulnerability-based targeting, and confirm none of your models or integrated third-party components fall within (a), (b), (f) or (g). Because Article 5 also covers placing on the market, shipping a banned capability to an EU customer is itself the violation, regardless of how the customer uses it. If you ship image, video or audio generation, add (ba) and (bb) to that audit before 2 December 2026: under Article 5(1a), point (a)(ii), safeguards must both prevent the output and correct "observed or reported" misuse.

If you're a deployer: Article 5 binds use, so you carry direct liability for operating a banned system even if a vendor built it. Emotion-recognition tools in recruitment or workforce monitoring are the most common trap. Contractual assurances from a vendor do not transfer the risk — the prohibition attaches to your use. Before you switch anything on, check it against Article 5 and, separately, confirm whether it also falls into a high-risk category under Annex III. Complipath's guided risk classification shows which of your systems face the high-risk deadlines of 2 December 2027 and 2 August 2028, with the reasoning on record.

Does anything you run fall under Article 5?

Classify your system now — 7 questions on the main line, plus follow-ups where they apply, no account, and the classification runs in your browser: answers stay there unless you choose to keep the result.

FAQ

What is the deadline for AI Act prohibited practices? Two deadlines now. The original eight prohibitions applied from 2 February 2025 under Article 113, third paragraph, point (a) — the Regulation's first binding deadline. The two added by Regulation (EU) 2026/1744, points (ba) and (bb), apply from 2 December 2026 under the exception written into that same point.

Are prohibited practices "high-risk" under the AI Act? No. Prohibited and high-risk are different tiers. High-risk systems (Annex III) are permitted subject to strict requirements. Article 5 practices are banned outright — no conformity assessment, registration or safeguard makes them lawful. Confusing the two is a common and costly error.

Is emotion recognition banned everywhere under the AI Act? No. Article 5(1)(f) prohibits inferring emotions only in the workplace and education institutions, and even there exempts medical and safety uses. Emotion recognition in other contexts may still be lawful but can trigger transparency duties under Article 50 or high-risk obligations depending on use.

What's the penalty for breaching Article 5? Up to €35 million or, "if the offender is an undertaking", 7% of total worldwide annual turnover for the preceding financial year, whichever is higher (Article 99(3)) — but for SMEs, including start-ups, Article 99(6) makes it whichever is lower. This is the Act's highest penalty tier. Member States set and enforce the fines; Chapter XII applied from 2 August 2025.

Sources: Regulation (EU) 2024/1689 (EUR-Lex), Articles 5, 99 and 113, as amended by Regulation (EU) 2026/1744 (EUR-Lex), Article 1, points (7) and (40). The European Commission's guidelines on prohibited practices are interpretive guidance, not binding law; where Article 5's boundaries turn on terms like "significant harm" or "materially distorting the behaviour of a person or a group of persons", the final interpretation rests with the Court of Justice. Our source corpus holds legislation and published Commission guidance, not case law — so nothing here should be read as a report on whether a court has ruled on these boundaries. Points (ba) and (bb) and the Article 5(1a) and (1b) qualifiers are new law that has not yet applied: no interpretation of terms such as "reasonably foreseeable and reproducible outcome" or "reasonable and adequate technical safety measures" is in our source corpus as of 12 August 2026, and our corpus does not include case law; verify before relying. The "without right" defence in point (bb) turns on national law, which differs between Member States and which we do not track.