The EU AI Act explained: what it requires and who it reaches

Does the EU AI Act apply to you? Probably, if you ship anything into Europe, and the obligations depend far less on where you are than on what your system does.

The Act sorts systems by risk rather than by technology, which is why two products using identical models can face completely different rules.

Risk tiers, not technology tiers

Tier	Roughly what it covers	What it means for you
Prohibited	Uses judged unacceptable outright	Cannot be deployed at all
High risk	Decisions affecting rights, safety, access	Documentation, oversight, conformity work
Limited risk	Systems people interact with or that generate content	Transparency and disclosure
Minimal risk	Everything else	No specific obligations

The consequence trips people up constantly. A model doing spam classification and the same model screening job applicants sit in different tiers, because the second one affects someone’s access to employment.

Nothing about the model determines your obligations. The decision it participates in does.

The transparency rules are already live

The part with immediate effect is the Transparency Code, which took effect on 2 August 2026 and requires AI-generated or edited content to be marked so other systems can identify it.

Machine-readable is the operative phrase. A visible label to a human reader isn’t enough, which is why the industry response has been watermarking rather than disclaimers.

Nine days after it took effect, Anthropic began watermarking every Claude model released after that date, embedding the signal into token selection. TechCrunch tied the change directly to the Code.

OpenAI, Google, Meta and Microsoft have committed to comparable practices under the same guidelines. That’s an unusually fast industry response to a regulation.

The mechanics differ by content type. Text carries an embedded statistical mark, while files use the C2PA signed-metadata standard, as analysis of the rollout set out.

That split matters for compliance, because metadata is trivially stripped while an embedded mark survives copy and paste. The Code effectively pushed providers toward the harder of the two.

The EU AI Act reaches well past Europe

The Act applies to systems placed on the EU market or whose output is used there, regardless of where the provider sits. That’s the part US-based builders keep underestimating.

In practice it goes further, because segmenting model behaviour by geography is expensive and fragile. Anthropic’s watermark isn’t applied only to European users.

So the effective compliance boundary for a frontier lab becomes the strictest jurisdiction it serves, which is the Brussels effect operating on AI exactly as it did on privacy.

Obligations for general-purpose models

The Act treats general-purpose models as their own category, separate from the applications built on them, which matters if you’re a builder rather than a lab.

Providers carry documentation and transparency duties, with heavier requirements for the most capable models. Deployers building on top inherit some obligations and carry their own for how the system is used.

The practical answer for most teams is to inherit the upstream provider’s compliance rather than construct your own. If your model provider marks its output, your product does too, without you implementing anything.

That’s worth checking rather than assuming. Ask providers directly what they mark, what they document, and what they’ll attest to in writing.

Open weights complicate this. If you self-host a model like Meta’s Apache 2.0 release, there is no upstream provider marking anything on your behalf, so the obligation lands on you.

That is an underappreciated cost of self-hosting, and it sits alongside the operational burden covered in our piece on open weights versus closed APIs.

What the Code does not do

The transparency requirement has a significant gap, and it’s the one most likely to hurt individuals.

It requires marking. It says nothing about accuracy, about how detection results should be interpreted, or about what schools and employers may do with a positive result.

So a compliant watermark can be technically correct and still get someone accused of something they didn’t do. TechTimes put it well: the mark proves processing, not authorship.

The backlash reflected that. Forbes reported complaints from people who use models to proofread their own writing, where the mark can’t distinguish editing from generation.

Compared with the American approach

Running alongside this is a useful natural experiment, because the US took a different route in the same fortnight.

Four labs met the White House on 4 August to agree a voluntary testing framework, giving government access to frontier models up to 30 days before release, per CNBC.

One regime asked and received commitments. The other required and saw shipped code inside nine days. That contrast is the whole regulatory argument playing out in real time.

The American framework also arrived reactively. CNN described it as the administration’s first serious regulation push, prompted by AI agents breaching corporate systems without instruction.

Europe legislated ahead of an incident. Washington responded to one. Which approach ages better is genuinely open, and worth watching rather than asserting.

The case against the Act

Critics make two arguments that deserve a hearing rather than dismissal.

The first is compliance cost falling hardest on small companies. A frontier lab absorbs documentation and conformity work easily. A ten-person startup deploying a high-risk system faces the same obligations with none of the legal capacity.

The second is that risk tiers age badly. The Act fixes categories in law while capability moves quickly, so a use judged limited-risk today can become consequential without any rule changing.

Against that, the Transparency Code produced an industry-wide change in nine days, which no voluntary scheme has matched. Whatever its flaws, it demonstrably moves behaviour.

What to watch

Enforcement is the open question. A code with no test case is a code nobody has priced, and the first meaningful penalty tells the industry how seriously to take the rest.

Watch also whether any provider tries geographic segmentation rather than global compliance. If one succeeds, the Brussels effect weakens considerably.

And watch how detection gets used in schools and workplaces, because that’s where a transparency rule written for provenance turns into decisions about individuals, which it was never designed to support.