A practical breakdown

How the EU AI Act works

The world's first comprehensive law for artificial intelligence. The trick to understanding it: it doesn't regulate the technology — it regulates what you do with it, and the rules scale with the risk.

2024In force since 1 Aug
4Risk tiers
€35MTop fine, or 7% of turnover
2025–28Phasing in

THE MENTAL MODEL

The Act is a sorting machine. Every AI system gets dropped in by its risk, and the weight of the rules matches the tier it lands in.

Most people expect a law that says "here are the rules for AI." The EU AI Act is built differently. It asks one question about each system — how much risk does this particular use create for people's rights and safety? — and sorts it into one of four tiers. Unacceptable uses are banned outright. High-risk uses carry heavy obligations. Limited-risk uses just need to be transparent. Everything else is free.

This is why the same piece of technology can sit in totally different tiers. A face-matching model is minimal-risk when it unlocks your phone, high-risk when it screens job applicants, and outright banned when it scrapes the open web to build a surveillance database.

Same tech · different use · different tier

Face recognitionunlocking your phoneMinimal
Face recognitionscreening job applicantsHigh-risk
Face recognitionscraping the web for a DBBanned
A chatbotanswering FAQsLimited

01 — THE FOUR TIERS

The risk pyramid

The fewer the systems in a tier, the heavier its rules. A tiny number of uses are banned at the top; the vast majority of everyday AI sits unregulated at the bottom. Click any tier to open it.

↑ Fewer systems · heavier rulesMore systems · lighter rules ↓

One more track, running alongside

General-purpose AI models — the large models like GPT, Gemini or Llama that power many products — are governed by their own separate set of rules, not the pyramid. We cover those further down.

02 — TRY IT YOURSELF

Which tier does it land in?

The whole Act turns on classification. Pick a real-world AI use below and see where the sorting machine drops it — and why.

Interactive · the classifier

← Pick a use case to classify it.

03 — UNACCEPTABLE RISK

The eight banned uses

At the top of the pyramid, a short list of practices is prohibited outright across the EU — no compliance path, no exceptions for most. These are uses seen as a clear threat to people's rights. (Article 5.)

The narrow exception

Real-time facial recognition in public by police is banned by default, but narrowly allowed for things like searching for a missing child or preventing an imminent terror attack — and only with a fundamental-rights assessment and judicial sign-off.

04 — HIGH RISK

High-risk: where most of the law lives

These systems are allowed, but heavily regulated — the bulk of the Act is about them. A use is high-risk if it's a safety component of a regulated product, or if it falls into one of eight sensitive areas (Annex III). Filter the areas, then see what providers must actually do.

IF IT'S HIGH-RISK, THE PROVIDER MUST…

05 — GENERAL-PURPOSE AI

The parallel track: general-purpose models

The big foundation models don't fit the use-based pyramid — one model can be put to a thousand uses. So they get their own rules, in two layers: a baseline for every GPAI model, plus extra duties for the most powerful ones that pose "systemic risk."

Every GPAI model

  • Keep technical documentation of how the model was built, trained and tested.
  • Give downstream developers the info they need to build on it compliantly.
  • Have a policy to respect EU copyright law.
  • Publish a summary of the training data used.

Free & open-licence models only owe the last two — unless they hit systemic risk.

+ systemic risk

  • Run model evaluations and adversarial testing (red-teaming).
  • Assess and mitigate systemic risks at their source.
  • Track and report serious incidents to the EU AI Office.
  • Ensure strong cybersecurity for the model and its weights.

These stack on top of the baseline — systemic-risk models owe both sets.

WHAT COUNTS AS "SYSTEMIC RISK"? — DRAG TO EXPLORE

The headline trigger is raw training compute. A model is presumed to pose systemic risk once the compute used to train it crosses 1025 floating-point operations (FLOP).

10²⁵FLOP of training compute
10²²10²³10²⁴10²⁵10²⁶10²⁷
Below the threshold — baseline GPAI rules apply, but not the systemic-risk duties.

The compute figure is the legal presumption; the Commission can also designate a model systemic on other grounds. Threshold illustrative of the rule, not a precise model ranking.

06 — WHO'S ON THE HOOK

Providers, deployers — and why location barely matters

The Act splits responsibility by role. Most of the weight falls on whoever builds the system, but the organisation that uses it has duties too.

Provider

The one who builds it

Develops the AI system (or has it built) and puts it on the market under their name. Carries the bulk of the obligations — risk management, documentation, testing, conformity assessment, registration.

Deployer

The one who uses it

Uses an AI system in a professional capacity (a hospital, a bank, an HR team). Lighter duties: use it as instructed, keep human oversight, monitor it, and — for some uses — inform the people affected.

It reaches beyond Europe's borders

Like the GDPR, the Act applies extraterritorially. If your AI system's output is used in the EU — even if your company sits in the US or anywhere else — you're in scope. That's why a law passed in Brussels shapes how AI is built worldwide.

07 — PENALTIES

What it costs to get it wrong

Fines are tiered to match the severity of the breach, and — crucially — calculated as the higher of a fixed sum or a percentage of global annual turnover. For large companies, the percentage is what bites.

€35M
or 7% of worldwide turnover
Using a banned practice
The most serious breach — deploying one of the prohibited AI uses.
€15M
or 3% of worldwide turnover
Breaking high-risk rules
Most other obligation breaches — by providers or deployers of high-risk systems.
€7.5M
or 1% of worldwide turnover
Giving false information
Supplying incorrect or misleading information to authorities.

DO THE MATH — THE "WHICHEVER IS HIGHER" CATCH

For SMEs and start-ups, the rule flips to whichever is lower, to keep fines proportionate. Figures are statutory caps, not automatic penalties.

08 — THE ROLLOUT

It arrives in waves, not all at once

The Act entered into force in August 2024, but its rules switch on in stages over several years — banned uses first, the heaviest high-risk duties last.

A moving target. In May 2026, EU institutions reached a provisional "Digital Omnibus" agreement to postpone the heaviest high-risk deadlines (Annex III high-risk to Dec 2027; product-embedded high-risk to Aug 2028) and simplify some rules. As of this writing it is agreed but not yet formally adopted, so dates below marked "proposed deferral" may still shift. Always check the official timeline before relying on a date.

09 — TEST YOURSELF

Did it click?

Five quick questions. You get instant feedback and the why behind each answer.

SOURCES

Where this comes from

This is an educational explainer, not legal advice. The EU AI Act is long and evolving; specific obligations depend on facts and on guidance still being issued. Consult the official text and qualified counsel for compliance decisions.