Article 5: Eight Bans, Already in Force, and Where the Lines Actually Sit

Most of the EU AI Act is about how to use AI responsibly. Article 5 is about what you may not do at all. Its eight prohibitions are absolute bans, with no conformity assessment, no documentation, no technical mitigation that makes a prohibited practice acceptable. And they have been directly applicable since 2 February 2025, so any organisation that has not screened its deployments against them is already exposed.

The bans are not all equally clear. Three carry most of the interpretive difficulty. The prohibition on manipulative and deceptive techniques in Article 5(1)(a) turns on subtle questions of intent and effect. The social-scoring ban in 5(1)(c) reaches further than people expect. And the limits on real-time remote biometric identification in 5(1)(h) are hedged with narrow, conditional exceptions that are easy to misread.

Social scoring deserves particular attention, because the boundary moved. The 2021 proposal limited the ban to public authorities; the final text removed that limitation. Large-scale private-sector behavioural scoring that produces unjustified or disproportionate detrimental treatment can now fall within the prohibition, which pulls in practices some firms regard as ordinary analytics.

The liability standard is the part that surprises people most. Article 5 uses an objective-or-effect formulation, which means a deployer can be caught even without intending a prohibited outcome, if that outcome was reasonably foreseeable. Foreseeability, not intention, is the test, and that materially widens the exposure.

The stakes match the seriousness. Article 5 infringements sit at the top penalty tier, up to 35 million euro or 7% of total worldwide annual turnover per infringement, whichever is higher. This is not the tier where you negotiate a remediation plan.

The practical response is a screening exercise, not a documentation exercise. Inventory the AI in use, test each system against the eight prohibitions, pay special attention to manipulation, scoring, and biometric identification, and apply the objective-or-effect standard rather than asking only what you intended.

Our whitepaper, Drawing the Line, takes each of the eight prohibitions in turn, examines the boundary issues, and provides a quick-reference and a timeline. Because there is no compliance pathway here, the only safe position is to know exactly where the lines are, and to stay well clear of them.