Comparison

High-risk vs prohibited AI

Written with AI from the sources below and checked against our facts sheet. Last checked 7 Nov 2026. How this site is made · Spot an error?

Part of: EU AI Act

The one-line rule

Prohibited AI (Art. 5) is banned outright, whatever safeguards you add. High-risk AI (Art. 6, Annexes I and III) is allowed but must meet strict requirements before and after it reaches the market.

Side by side

	Prohibited AI	High-risk AI
Legal basis	Art. 5	Art. 6, with Annex I (products) or Annex III (uses)
Status	Banned	Allowed, with strict requirements
Can safeguards make it lawful?	No; consent and CE marks do not help	Yes; meeting the requirements is the point
Examples	Social scoring, untargeted facial scraping, workplace emotion recognition	Hiring tools, credit scoring, exam proctoring, AI in medical devices
Who owes duties	Anyone who places, puts into service or uses the practice	Mainly providers; also deployers, importers and distributors
Exceptions	Narrow and listed, such as medical or safety reasons for emotion recognition	The Art. 6(3) exit for narrow tasks, never for profiling
Applies from	2 Feb 2025; two Omnibus bans from 2 Dec 2026	Annex III from 2 Dec 2027; Annex I from 2 Aug 2028
Maximum fine	€35 million or 7% (Art. 99(3))	€15 million or 3% (Art. 99(4))
Registration	Not applicable	Annex III systems go in the EU database (Art. 49)

When each applies

Start with Art. 5. If a practice is on the list, the analysis ends. No conformity assessment, consent form or human reviewer can make it lawful. The original eight bans have applied since 2 February 2025.

The Digital Omnibus added two bans in new points (ba) and (bb). The first covers AI that generates realistic intimate or sexually explicit images, video or audio of an identifiable person without explicit consent. The second covers AI that generates child sexual abuse material. Under new Art. 5(1a), a provider is caught if that is the intended purpose. It is also caught if such output is a reasonably foreseeable, reproducible result and the system lacks adequate safeguards. A deployer is caught when it uses a system for that purpose. These bans apply from 2 December 2026 (Art. 113, as amended). See deepfake.

Then check Art. 6. A system is high-risk through Annex I if it is a safety component of a regulated product, or is itself one. That product must also need third-party conformity assessment. A system is high-risk through Annex III if used in one of eight areas. They are biometrics, critical infrastructure, education, employment, essential services, law enforcement, migration, and justice and democracy. The Art. 6(3) exit covers narrow procedural or preparatory tasks, but a system that profiles people is always high-risk. Details are on the high-risk AI system page.

Near misses are where the exam lives. The same technology can land in different tiers:

Prohibited	High-risk instead
Emotion recognition at work or school (Art. 5(1)(f))	Emotion recognition elsewhere, such as in a store (Annex III 1(c))
Real-time remote biometric identification by police in public spaces (Art. 5(1)(h))	Remote biometric identification after the fact (Annex III 1(a))
Predicting a person's offending based solely on profiling (Art. 5(1)(d))	Police risk assessments not based solely on profiling (Annex III 6(d))
Biometric categorization to infer race, religion or sexual orientation (Art. 5(1)(g))	Biometric categorization by other sensitive or protected attributes (Annex III 1(b))

Worked example

Veltro Logistics, a fictional German warehouse operator, reviews three AI proposals.

Proposal one reads facial expressions on break-room cameras to flag "disengaged" staff for coaching. That is emotion recognition in the workplace, banned by Art. 5(1)(f). Staff consent does not lift the ban, and neither does the vendor's CE mark. Veltro drops it.

Proposal two allocates shifts based on each worker's productivity history. Task allocation based on individual behavior is an Annex III point 4(b) use, so it is high-risk. Veltro can proceed as a deployer from 2 December 2027. It must follow the instructions for use, assign trained human oversight and keep logs for at least six months. It must also inform workers' representatives and affected workers before use. See EU AI Act high-risk obligations by role.

Proposal three is a chatbot answering staff questions about the leave policy. It makes no employment decisions, so it is not high-risk. It falls under Art. 50: people must be told they are talking to an AI, unless that is obvious.

The exam trap

The classic trap is a near miss: a prohibited practice dressed up with safeguards, or a high-risk use described as banned.

The tells:

"Consent," "CE marked" or "a human reviews it" never rescues an Art. 5 practice.
"All emotion recognition is banned" is wrong. Only workplace and education uses are, with medical and safety exceptions.
"High-risk means banned" is wrong. High-risk means allowed with duties.
A human in the loop does not take a system out of high-risk.
Social scoring is banned for private actors too, not only governments.

Watch the dates. The original bans applied from 2 February 2025, the Omnibus bans from 2 December 2026. Annex III duties start on 2 December 2027. Drill Art. 5 in lesson II.C.2 and classification in lesson II.C.3.

Frequently asked questions

What AI is prohibited under the EU AI Act?

Art. 5 bans eight original practices: harmful manipulation, exploiting vulnerabilities, social scoring and crime prediction based solely on profiling. It also bans untargeted facial-image scraping, emotion recognition at work or school, and biometric categorization by sensitive traits. The eighth is real-time remote biometric identification by police in public spaces, with narrow exceptions. The 2026 Omnibus added bans on non-consensual intimate deepfakes and child sexual abuse material, from 2 December 2026.

Is emotion recognition banned in the EU?

Only in the workplace and in education, unless used for medical or safety reasons (Art. 5(1)(f)). Elsewhere, such as in retail, emotion recognition is high-risk under Annex III point 1(c). Deployers must also inform the people exposed to it (Art. 50(3)).

What makes an AI system high-risk under the EU AI Act?

There are two routes. Under Art. 6(1), AI that is a safety component of a product, or is itself a product, under Annex I laws and needs third-party conformity assessment. Under Art. 6(2), AI used in one of the eight Annex III areas, such as hiring or credit scoring. Art. 6(3) lets some Annex III systems out, but never those that profile people.

What are the fines for prohibited AI vs high-risk AI?

Prohibited practices carry up to €35 million or 7% of worldwide annual turnover, whichever is higher (Art. 99(3)). Breaching high-risk duties carries up to €15 million or 3% (Art. 99(4)). SMEs, including start-ups, face the lower of the two figures. Since the Omnibus, small mid-caps also get the lower figure, but only for the €15 million and €7.5 million tiers.

Sources

Spot an error? Tell us