The AI Act

He teaches a course called The Law of Artificial Intelligence, so this is his taught subject, and he has published on it twice in ways that matter to you. Both are summarised below so you do not need to chase them.

Regulation (EU) 2024/1689, adopted 13 June 2024, in force 1 August 2024. The first comprehensive horizontal AI regulation anywhere.

Its architecture is product safety law — conformity assessment, CE marking, notified bodies, market surveillance — with a fundamental rights layer added on top. That hybrid origin explains most of its awkwardness, and it is a fair criticism to voice: a framework designed for certifying lifts and toys is being asked to police fundamental rights.

The Omnibus also added a prohibition on generating non-consensual intimate imagery and child sexual abuse material, and moved the regulatory sandbox deadline to December 2027.

1. Unacceptable risk — prohibited outright, Article 5
2. High risk — permitted but heavily regulated, Article 6 with Annexes I and III
3. Limited risk — transparency obligations only, Article 50
4. Minimal risk — unregulated

Running alongside all four: a separate regime for general-purpose AI models, Articles 51 to 56, with additional duties above a systemic-risk threshold.

Caught: providers placing systems on the EU market wherever they are established; deployers established or located in the EU; and providers or deployers in third countries where the output is used in the EU.

Excluded: military, defence and national security purposes; scientific research and development as the sole purpose; pre-market research and development, though not real-world testing; purely personal non-professional use; and free and open-source AI, except where it is prohibited, high-risk, or caught by Article 50.

AI system, Article 3(1): a machine-based system, operating with varying levels of autonomy, possibly adaptive after deployment, that infers from input how to generate outputs such as predictions, content, recommendations or decisions influencing physical or virtual environments. Aligned with the OECD definition.

General-purpose AI model, Article 3(63): displays significant generality and competently performs a wide range of distinct tasks. Systemic risk is presumed under Article 51 where cumulative training compute exceeds ten to the power of twenty-five floating point operations.

• Subliminal, manipulative or deceptive techniques materially distorting behaviour and causing significant harm
• Exploiting vulnerabilities based on age, disability, or social and economic situation
• Social scoring by public or private actors leading to detrimental treatment that is unjustified or disproportionate
• Individual criminal risk assessment based solely on profiling or personality traits, that is, predictive policing
• Untargeted scraping of facial images to build recognition databases
• Emotion recognition in workplaces and education
• Biometric categorisation inferring race, political opinions, trade union membership, religion, sex life or sexual orientation
• Real-time remote biometric identification in public spaces for law enforcement, with narrow exceptions

Two of these nearly catch competition screening, and the near-miss is the interesting part.

Article 6(1): high-risk where the system is a safety component of, or is itself, a product under Annex I legislation requiring third-party conformity assessment.

Article 6(2): the Annex III use cases are high-risk.

Article 6(3) derogation: an Annex III system is not high-risk where it poses no significant risk — because it performs a narrow procedural task, improves a previously completed human activity, detects decision patterns without replacing or influencing human assessment, or performs preparatory work. But it is always high-risk if it profiles natural persons.

The eight Annex III areas: biometrics; critical infrastructure; education; employment and worker management; access to essential public and private services, which is where credit scoring sits; law enforcement; migration, asylum and border control; and administration of justice and democratic processes.

And an agency wanting to stay outside would reach for Article 6(3): our screen merely detects patterns without influencing human assessment. After SCHUFA, that is a factual claim a court will test — and if the agency opens an investigation because the screen flagged it, the claim looks thin.

• Article 16 — provider obligations: quality management, documentation, conformity assessment, CE marking, registration, corrective action
• Article 25 — the value chain, and when a deployer becomes a provider
• Article 26 — deployer obligations: use per instructions, assign human oversight to competent and trained people, ensure input data relevance, monitor, retain logs, cooperate with authorities. Article 26(11) requires deployers of Annex III high-risk systems making or assisting decisions about natural persons to inform those persons
• Article 27 — the fundamental rights impact assessment, required of deployers that are bodies governed by public law or private entities providing public services. It must describe the deployment process, the period of use, the categories of persons affected, the specific risks of harm, the human oversight measures, and the governance and complaint arrangements, and be notified to the market surveillance authority

A competition agency is a body governed by public law, so Article 27 would bite — except that it only triggers for Annex III high-risk systems. The gap from section 7 switches it off.

Systems interacting with humans must disclose that they are AI. Synthetic content must be marked in machine-readable form. Deep fakes must be disclosed. Text published to inform the public on matters of public interest must be labelled as AI-generated unless a human took editorial responsibility for it.

• Article 51 — systemic risk classification, presumed above ten to the twenty-fifth floating point operations of cumulative training compute
• Article 53 — provider duties: technical documentation, information to downstream providers, a copyright policy, and a public summary of training content
• Article 55 — systemic risk duties: model evaluation including adversarial testing, systemic risk assessment and mitigation, serious incident reporting, cybersecurity
• Article 56 — codes of practice
• Partial open-source exemption under Article 53(2), unless the model carries systemic risk

• Article 85 — the right to lodge a complaint with a market surveillance authority
• Article 86 — the right to explanation of individual decision-making. A person subject to a decision taken by a deployer on the basis of output from an Annex III high-risk system, producing legal effects or adversely affecting health, safety or fundamental rights, may obtain clear and meaningful explanations of the role of the AI system in the decision procedure and the main elements of the decision taken
• Article 87 — whistleblower protection

Small and medium enterprises pay the lower of the two figures rather than the higher.

• GDPR and Regulation 2018/1725 apply in parallel, under Article 2(7)
• Product Liability Directive (EU) 2024/2853 now treats software and AI as products
• The AI Liability Directive proposal was withdrawn in 2025, so there is no harmonised fault-based AI liability regime. Worth knowing: it is a visible hole in the EU's framework

• Product-safety architecture poorly suited to fundamental rights harms
• Annex III is a closed list and already dated. Your gap is an instance of this
• Article 6(3) hands providers substantial self-assessment discretion
• Heavy dependence on harmonised standards that are not finished
• Public-sector deployment gaps, and a weak penalty regime for public bodies
• Thin national enforcement capacity
• The Digital Omnibus deferral widely read as deregulation under industry pressure

Of the project's three strands — accuracy, fairness, and institutional arrangements — the AI Act is the main instrument for the second, and it is where the legal-track PhD on the fairness problem will spend four years.

Schrepel's own framing of that strand is worth quoting from memory: AI systems inherit the biases of their training data, the computer science literature on that point is now vast, and competition enforcement has barely engaged with it. When a screening model flags one market rather than another, the selection reflects choices about data and design that nobody outside the agency can examine. When machine-generated analysis informs a decision, the duty to give reasons meets an explainability problem the case law never anticipated. The AI Act adds obligations, but its application to enforcement agencies using AI against private parties raises questions the text does not settle.

1. Regulation (EU) 2024/1689, in force August 2024, product-safety architecture with a fundamental rights layer.
2. Annex III high-risk obligations now apply from 2 December 2027, deferred by the Digital Omnibus in July 2026. Most summaries are wrong about this.
3. Four tiers — prohibited, high-risk, transparency-only, minimal — plus a separate general-purpose AI regime.
4. Provider versus deployer decides your obligations, and Article 25 means a deployer can become a provider.
5. Article 5 prohibitions nearly catch competition screening through social scoring and predictive policing, but miss, because firms are not natural persons and competition enforcement is not criminal.
6. The Annex III gap: point 6 is criminal law enforcement, point 8 is judicial authorities, a competition agency is neither, and the gap cascades through Articles 10, 14, 27 and 86.
7. Article 10 requires representative data examined for bias; Article 14 requires meaningful oversight and names automation bias explicitly.
8. Article 86 gives a right to the system's role and the main elements of the decision — narrower than people assume, and only for Annex III high-risk.
9. Article 55 makes adversarial testing a legal obligation.
10. Penalties are 7, 3 and 1 percent — but Article 100 caps EU institutions at roughly 1.5 million euro. Asymmetric.