The world's first comprehensive AI law, mapped for the classroom.
The EU AI Act doesn't regulate AI in general — it regulates AI by the risk a specific use poses. The same model can be unregulated in one use and heavily regulated in another. That single idea is the whole framework.
Fewer systems at the top, stricter rules. Most AI in use today falls in the bottom two tiers.
Purpose & Origin
What the Act is, why the EU built it this way, and who it actually binds.
What it is
Regulation (EU) 2024/1689 — the "AI Act" — is the first comprehensive, horizontal legal framework for AI adopted by a major regulator. As an EU Regulation, it applies directly in all member states without needing separate national implementing legislation, unlike a Directive.
Why it was built as a risk pyramid
Rather than defining "AI" and regulating it uniformly, the Act sorts AI systems by the risk a specific use poses to health, safety, and fundamental rights. A face-recognition model used to unlock a phone and one used for mass surveillance are the same technology facing entirely different rules — because the Act regulates the use, not the algorithm.
Who it binds
It has extraterritorial reach similar to GDPR: it applies to providers placing AI systems on the EU market and to providers and deployers outside the EU if the system's output is used within the EU — meaning a U.S. company can fall under it without any EU office.
What's explicitly excluded
Military and defense AI, AI used solely for scientific research and development, and systems not yet placed on the market are outside scope, along with some exemptions for open-source models that don't fall into higher-risk categories.
How to Use This Guide
Built for a course session — not a substitute for legal advice or a compliance program.
- Classify before anything else. Every question about the Act reduces to "which tier is this system in?" Get comfortable sorting examples into unacceptable / high / limited / minimal before studying the obligations attached to each.
- Notice the pyramid is about volume too. A handful of practices are banned outright (top). A defined but real list of high-risk use cases carries heavy obligations. Most deployed AI — recommendation engines, spam filters, most generative-AI chat tools — sits in the bottom two tiers.
- Separate "the system" from "the model." Chapter V's General-Purpose AI rules are a second, mostly separate track — they attach to foundation models themselves (like an LLM), independent of what tier a specific downstream application built on that model lands in.
- Track the phased timeline. Almost every debate about "is this legal yet" comes down to which obligations have actually taken effect versus which are legislated for a later date.
- Go to EUR-Lex for anything graded. This guide paraphrases; official citations should reference the Regulation's actual article and recital numbers from the consolidated EUR-Lex text.
Phased Timeline
The Act didn't take full effect on day one — obligations phase in over roughly three years from entry into force.
Entry into force
The Regulation is published and enters into force EU-wide, starting the clock on every later deadline.
Prohibitions & AI literacy take effect
Article 5 banned practices become enforceable, along with Article 4's requirement that providers and deployers ensure staff have adequate AI literacy.
GPAI rules, governance & penalties take effect
Chapter V obligations for general-purpose AI models apply; the AI Office and national competent authorities must be operational; the penalty regime becomes enforceable.
Main body of the Act applies
Most high-risk AI system obligations (Annex III use cases) and Article 50 transparency obligations become enforceable — the deadline this guide's status banner flags as imminent.
Remaining high-risk obligations apply
High-risk AI systems that are safety components of products already regulated under EU product-safety law (Annex I) come fully under the Act's obligations.
Unacceptable Risk — Article 5
A short, specific list of practices considered such a clear threat to fundamental rights that they're prohibited outright — no risk assessment or mitigation makes them acceptable.
- Subliminal, manipulative, or deceptive techniques that materially distort behavior and cause harm.
- Exploiting vulnerabilities of specific groups (age, disability, socio-economic situation) to distort their behavior in a harmful way.
- Social scoring by public authorities based on behavior or characteristics, leading to detrimental treatment unrelated to the context in which the data was generated.
- Individual crime-risk assessment based solely on profiling or personality traits, without objective, verifiable facts directly linked to criminal activity.
- Untargeted scraping of facial images from the internet or CCTV to build facial-recognition databases.
- Emotion inference in the workplace and education, except for narrow medical or safety reasons.
- Biometric categorization inferring sensitive attributes (race, political opinion, sexual orientation, religion) from biometric data.
- Real-time remote biometric identification by law enforcement in publicly accessible spaces — allowed only under narrowly defined, judicially authorized exceptions (e.g., searching for a specific victim of a serious crime).
High-Risk AI Systems — Articles 6–15
Legal, but only with substantial obligations. This is where most of the Act's compliance weight sits — a defined list of use cases (Annex III) plus AI used as a safety component in already-regulated products (Annex I).
Biometrics
Remote biometric identification, categorization, emotion recognition (where not outright banned).
Critical infrastructure
Safety components managing critical digital infrastructure, road traffic, water, gas, heating, electricity.
Education & training
Determining access, admission, or assessing/scoring students — the category most directly relevant to your own institution.
Employment
Recruitment, screening applications, promotion/termination decisions, task allocation, performance monitoring.
Essential services
Eligibility for public benefits, creditworthiness scoring, life/health insurance pricing, emergency dispatch prioritization.
Law enforcement
Risk assessment, polygraph-style tools, evidence evaluation, crime-analytics (subject to narrower carve-outs than Art. 5).
Migration, asylum, border control
Risk and security assessments, examination of asylum/visa applications.
Justice & democratic processes
Assisting judicial research/interpretation, and AI intended to influence elections or referenda through direct interaction with voters.
- Risk management system spanning the whole lifecycle, continually updated.
- Data governance — training, validation, and testing data must meet quality criteria and be examined for bias.
- Technical documentation & record-keeping (automatic logging) sufficient to demonstrate compliance.
- Transparency & instructions for use clear enough for deployers to use the system properly and interpret its output.
- Human oversight designed in, so a natural person can effectively oversee the system's operation, including the ability to intervene or stop it.
- Accuracy, robustness & cybersecurity appropriate to the intended purpose, throughout the system's lifecycle.
Organizations that use a high-risk system (not just build it) must use it per the provider's instructions, assign human oversight to competent people, monitor its operation, and — for certain public-sector and high-impact private uses — conduct a fundamental rights impact assessment before deployment.
Limited Risk — Transparency Obligations (Article 50)
These systems aren't restricted in what they can do — they just have to tell people what they're interacting with.
- Chatbots & conversational AI. Users must be informed they're interacting with an AI system, unless it's obvious from context.
- Synthetic media / deepfakes. AI-generated or manipulated image, audio, or video content resembling real people, places, or events must be labeled as artificially generated or manipulated.
- Emotion recognition & biometric categorization systems (where not banned under Art. 5) must inform the people exposed to them.
- AI-generated text published to inform the public on matters of public interest must be disclosed as AI-generated, with narrow exceptions (e.g., human editorial review with accountability).
Minimal / No Risk
The tier holding most AI in everyday use — spam filters, recommendation systems in non-high-risk contexts, AI-enabled video games, inventory-management tools.
No mandatory obligations under the Act. Providers are encouraged (not required) to voluntarily adopt codes of conduct reflecting the high-risk requirements, on a best-effort basis. General EU law — GDPR, consumer protection, product liability — still applies regardless of AI Act tier.
General-Purpose AI Models — Chapter V
Rules that attach to foundation models themselves (like large language models), largely independent of the risk-tier pyramid above, which applies to specific AI systems and their use.
Baseline obligations — all GPAI providers
Maintain technical documentation, provide information to downstream providers integrating the model, publish a policy for complying with EU copyright law, and publish a sufficiently detailed summary of the content used to train the model.
Additional obligations — "systemic risk" models
Models trained with cumulative compute above a defined threshold (10^25 floating-point operations) are presumed to carry systemic risk, triggering model evaluation, adversarial testing, systemic-risk mitigation, incident tracking and reporting, and cybersecurity protections.
Governance & Penalties
Who enforces the Act, and what non-compliance actually costs.
EU AI Office
A body within the European Commission overseeing GPAI models directly and coordinating enforcement across member states.
National competent authorities
Each member state designates authorities responsible for market surveillance and enforcement within its territory for most non-GPAI matters.
| Violation | Maximum penalty |
|---|---|
| Article 5 prohibited practices | Up to €35M or 7% of global annual turnover, whichever is higher |
| Other obligations (e.g., high-risk requirements) | Up to €15M or 3% of global annual turnover, whichever is higher |
| Supplying incorrect, incomplete, or misleading information to authorities | Up to €7.5M or 1% of global annual turnover, whichever is higher |
Classroom Uses
Ways instructors have used this structure with students.
Tier-sorting exercise
Give students a list of real AI products (a resume screener, a grammar checker, a facial-unlock feature, a predictive-policing tool) and have them sort each into a tier with a one-sentence justification.
"System vs. model" untangling
Have students take a single LLM-based product and separately identify what GPAI obligations apply to the underlying model versus what risk-tier obligations apply to the specific application built on it.
Deployer vs. provider debate
Assign half the class "provider" and half "deployer" for a high-risk education-sector use case, and have each side identify what it owes the other under Articles 16 and 26.
Timeline current-events check
Have students search for the current status of the August 2026 milestone and report back — a live lesson in how quickly regulatory timelines can shift.
Sources & Further Reading
This guide is a teaching summary — verify current dates and obligations before using in graded work.
EUR-Lex — official consolidated text
The authoritative, article-numbered legal text.eur-lex.europa.eu — search "Regulation (EU) 2024/1689"
European Commission — AI Act page
Official implementation timeline, guidance documents, and news on any simplification/delay proposals.digital-strategy.ec.europa.eu/en/policies/regulatory-framework-ai
EU AI Office
Guidance specific to general-purpose AI model obligations and the Code of Practice for GPAI providers.digital-strategy.ec.europa.eu/en/policies/ai-office