Purpose
Art. 5 AI Act (Regulation (EU) 2024/1689) lists the AI practices that are regarded as incompatible with European fundamental rights fully prohibited are. These prohibitions have applied since 2 February 2025. Infringements constitute the highest fine tier of the Regulation. This skill reviews whether a system fulfils a prohibition offence and whether a narrowly construed exception applies.
⚠️ Currency (as of 2026-07): Two new prohibition offences from 02.12.2026. The Digital Omnibus on AI (European Parliament 16.06.2026, Council 29.06.2026) has expanded the catalogue of Art. 5 AI Act by two practices that are from 02.12.2026 prohibited:
New prohibition offence Applicability Non-consensual intimate imagery (NCII) — AI-supported creation/dissemination of intimate depictions without the consent of the person depicted 02.12.2026 Child sexual abuse material (CSAM) 02.12.2026 The existing prohibitions of Art. 5 continue to apply unchanged since 02.02.2025 and are neither postponed nor substantively amended by the Digital Omnibus neither postponed nor substantively amended. The new provisions are added alongside the existing catalogue; the lettering of the existing offences remains in place. Official Journal reference of the amending regulation
[unverifiziert - prüfen].Supervision expanded: The competence of the AI Office now also extends to AIsystemsthat are based on GPAI models which were developed within the same undertaking as the model provider. Applications built within the group therefore no longer fall solely under national market surveillance.
National competence in Germany: The market surveillance authority is the Federal Network Agency (residual competence under the KI-MIG), not the data protection supervisory authority or the BfDI. For biometric systems in law enforcement, border control, the judiciary and democratic processes, the independent AI Market Surveillance Chamber (UKIM) is competent. The KI-MIG passed the Bundestag on 11.06.2026 and the Bundesrat on 10.07.2026 passed, but as of 21.07.2026 had not yet been promulgated in the BGBl — a BGBl reference therefore does not exist and must not be cited.
[unverifiziert - prüfen]
Inputs
- Function and purpose of the AI system
- Addressees (general public, employees, vulnerable groups)
- Deployment context (publicly accessible space, law enforcement, private, governmental)
- Techniques used (manipulation, biometric capture, assessment of social behaviour)
- Role of the client (provider, deployer)
Sub-agent architecture
Three conceptual roles examine the system. A fact-pattern assessor assigns the function to the eight prohibited categories of Art. 5 of the AI Act and works out the respective constituent elements. An exception analyst conversely examines the narrowly defined exceptions, in particular for real-time remote biometric identification for law enforcement purposes. A penalty assessor quantifies the fine risk under Art. 99 of the AI Act and distinguishes it from high-risk classification. The fact-pattern assessor passes every affirmed prohibited practice to the exception analyst before the final result is established.
Process
1. Prohibited categories (Art. 5(1) of the AI Act)
a) Subliminal/manipulative influencethat materially changes behaviour and causes harm b) Exploitation of vulnerability (age, disability, social/economic situation) with harmful consequences c) Social scoring — assessment/classification based on social behaviour or personal characteristics with unjustified detrimental treatment in unrelated contexts d) Predictive policing based solely on profiling/personality traits e) Untargeted scraping of facial images to build facial recognition databases f) Emotion recognition in the workplace and in educational institutions (except for medical/safety-related purposes) g) Biometric categorisation to infer sensitive characteristics (e.g. ethnic origin, political opinion, sexual orientation) h) Real-time remote biometric identification in publicly accessible spaces for law enforcement purposes
The practices a) to h) have applied since 2 February 2025 and remain unchanged.
1a. New prohibited practices from 2 December 2026 (Digital Omnibus on AI)
In addition — and only from 2 December 2026 — the following are prohibited:
- Non-consensual intimate imagery (NCII). This covers AI systems that generate or manipulate intimate or sexualised depictions of real persons without their consent ("nudify" applications, deepfake pornography). This affects providers of generative image and video models with the relevant functionality as well as deployers who make such systems available.
- Child sexual abuse material (CSAM). AI-supported generation of such content is prohibited regardless of whether real persons are depicted.
Review notes:
- Observe temporal application. Before 2 December 2026, these practices are not prohibited under Art. 5 of the AI Act. However, they are already regularly subject to national criminal and personality rights inadmissible (including §§ 184b, 201a StGB, general right of personality, GDPR) — the statement "currently still permitted" is in the mandate false.
- Advance obligation of the providers. Until the deadline, generative models must be safeguarded by filters, prompt and output controls as well as abuse reporting channels; the conversion cannot be achieved at short notice.
- Distinction from Art. 50. The labelling obligation for deepfakes (Art. 50(4), from 02.08.2026) does not make prohibited content permissible — a labelled NCII depiction remains prohibited from 02.12.2026.
Reference to the amending regulation in the Official Journal [unverifiziert - prüfen].
2. Exceptions for real-time remote biometric identification (Art. 5(1)(h) AI Act)
Only three narrowly defined purposes come into consideration:
- targeted search for victims of kidnapping, human trafficking, sexual exploitation as well as missing persons
- averting a specific, substantial and imminent danger to life or limb or a terrorist attack
- localisation/identification of a person suspected of a serious criminal offence
Requirement: prior judicial or administrative authorisation (Art. 5(3) AI Act), proportionality, national legal basis.
3. Distinction from high-risk classification
If no prohibition applies, the high-risk assessment under Art. 6 in conjunction with Annex III AI Act is the next step — a system can be permissible but highly regulated.
4. Sanctions (Art. 99 AI Act)
Violations of the prohibitions under Art. 5 AI Act are punished with fines of up to EUR 35 million or 7% of worldwide annual turnover — the highest tier.
Risks
- Manipulation — seemingly harmless personalisation can be prohibited as subliminal influence under Art. 5 AI Act.
- Social scoring through the back door — overarching reputation or creditworthiness systems can fulfil the prohibition.
- Apparent exception — the exceptions for real-time remote biometric identification are narrow; a missing authorisation under Art. 5(3) AI Act makes the use prohibited.
- Fine risk — up to EUR 35 million / 7% under Art. 99 AI Act; the prohibitions have already applied since February 2025.
- Overlooking the new prohibitions on NCII and CSAM — the catalogue in Art. 5 has been expanded by the Digital Omnibus on AI; the two offences apply from 02.12.2026. Reviews that only work through the eight offences from 2025 are incomplete. Providers of generative image and video models are particularly affected.
- "Permitted until 02.12.2026" misunderstood as a licence — the generation of NCII and CSAM is already subject to sanctions today under national criminal and personality law. The AI Act merely adds another, particularly severe sanction level.
- National jurisdiction wrongly addressed — The market surveillance authority is the Bundesnetzagentur (fallback competence, KI-MIG), not the data protection supervisory authority or the BfDI; for biometrics in law enforcement, border control, the judiciary and democratic processes, the independent KI-Marktüberwachungskammer (UKIM). As of 21.07.2026, the KI-MIG had not yet been promulgated in the BGBl.; a BGBl. citation must not be given.
[unverifiziert - prüfen] - Intra-group GPAI applications incorrectly classified — if an AI system is based on a GPAI model of the same company, the supervision of the AI Office is opened; the assumption of purely national competence does not hold.
Output format
VERBOTENE KI-PRAKTIKEN — PRÜFUNG — <System> — <Datum>
I. Funktion / Zweck [Beschreibung]
II. Einschlägige Kategorie (Art. 5 KI-VO) [lit. a–h / NCII / CSAM / keine]
Tatbestandsmerkmale [erfüllt? <…>]
Geltung ab [02.02.2025 / 02.12.2026]
III. Ausnahme (Art. 5 Abs. 1 lit. h / Abs. 3 KI-VO) [greift? / Genehmigung?]
IV. Ergebnis [VERBOTEN / nicht verboten]
V. Falls nicht verboten: Hochrisiko-Check (Art. 6, Anhang III) [erforderlich?]
VI. Bußgeldrisiko (Art. 99 KI-VO) [bis 35 Mio. EUR / 7 %]
Empfehlung: <…>
Nächste Schritte: <…>
Sources
Statute
- Regulation (EU) 2024/1689 (AI Regulation / EU AI Act) — full text with annexes
- Art. 5 AI Regulation (prohibited practices), Art. 6 AI Regulation and Annex III (high-risk), Art. 99 AI Regulation (sanctions)
- Digital Omnibus on AI — amending regulation to the AI Regulation (EP 16.06.2026, Council 29.06.2026); new prohibitions NCII and CSAM from 02.12.2026. Official Journal citation
[unverifiziert - prüfen] - KI-MIG (AI Market Surveillance and Innovation Promotion Act) — Bundestag 11.06.2026, Bundesrat 10.07.2026, not promulgated in the BGBl. as of 21.07.2026; competence BNetzA / UKIM
[unverifiziert - prüfen]
Guidelines
- European Commission, Guidelines on the application of the prohibitions pursuant to Art. 5 AI Regulation (February 2025) — do not yet cover the NCII and CSAM offences
Secondary literature
- Hilgendorf/Roth-Isigkeit, KI-VO, 1st ed. 2024
- BeckOK KI-VO (Online)