Skip to main content

Europe Wrote the AI Rulebook. The Hard Part Is Getting Others to Follow It

Ilustrační obrázek
A company in Toronto or Seoul that wants to sell an AI hiring tool in Europe now has to answer questions its engineers would not hear at home: where the training data came from, whether the output can be marked as machine-generated, and who signed off on the risk assessment. None of that is a side effect of the EU AI Act. It is the mechanism.

That mechanism is the subject of a study published in Information Systems Frontiers on 31 October and covered by Bioengineer.org on 21 November. Its author, Vasiliki Koniakou, asks a question that matters to anyone selling software into the single market: can a regulator with no frontier AI lab of its own still decide how AI gets built elsewhere?

The paper's answer is careful. Europe has already won the argument about vocabulary. Whether it wins the argument about practice is a different question, and one the text leaves open.

Three models of power

Koniakou sets the EU against two rivals with very different strengths. The United States leads through capability: compute, venture capital, and a handful of platforms large enough to set their own technical defaults. China leads through coordination: state direction, national champions, and rules written to be executed rather than negotiated.

The EU's model is leadership through governance. Brussels writes the rule, opens a market of roughly 450 million consumers, and waits for the compliance cost to travel upstream. A developer who wants European customers redesigns the product for Europe, then often ships that same design everywhere else. Anu Bradford called this the Brussels Effect, and the AI Act is its largest test so far in software.

What the paper actually measures

The study is not a single number. It builds a framework with four dimensions for judging the strategy: strategic intent, anticipatory rulemaking, governance instrumentation, and external uptake. It then lists four conditions that decide whether the strategy travels: structural power, institutional capacity, relational receptivity, and strategic durability. Three mechanisms are examined as routes to advantage: harmonising the internal market, building trust that drives adoption, and differentiating European products by their compliance record.

Read plainly, that means Brussels has the market size and the legal text, and now needs standards bodies, national supervisors, and foreign regulators willing to copy. The first two are its own to fix. The third is not.

What Europe has already achieved

The vocabulary is the clearest win. Terms like general-purpose AI, high-risk system and systemic risk now appear in draft laws and corporate policy documents well outside the EU. Risk tiers, which sound bureaucratic in a Commission press release, turned out to be a portable way of talking about harm.

The legal machinery behind the words started running this year. Transparency duties under Article 50 of the AI Act took effect in August 2026. In practice they require disclosure when someone is interacting with an AI system, unless that is obvious. For synthetic media, the duties split. Under Article 50(2), providers of generative AI systems must mark outputs in a machine-readable format. Under Article 50(4), a deployer who publishes synthetic images or deepfakes has a disclosure duty. A freelance illustrator in Brno who publishes generated visuals on a client's website is a deployer, not the model provider: the Article 50(4) disclosure duty applies to them, while the provider must meet the Article 50(2) marking duty.

Penalties, in numbers people can picture

For the most serious breaches, Article 99(3) sets the ceiling at €35 million or 7% of worldwide annual turnover, whichever is higher. The GDPR, by comparison, tops out at €20 million or 4%. The “whichever is higher” matters: a €10 million-revenue undertaking that is not an SME faces a ceiling of €35 million, not €700,000.

Article 99(6) changes the arithmetic for SMEs and start-ups. It caps their fine at the lower of the two amounts. That is why a €10 million-revenue SME tops out at roughly €700,000 — 7% of turnover. For that SME, the figure is not theoretical: it would erase the year.

Penalties are not the same as enforcement capacity, though. Fines only work if supervisors have staff who can read a technical file and tell whether a training data summary is real. That capacity is uneven across member states, and the paper treats institutional strength as one of its four enabling conditions rather than a given.

An unfinished timetable

The obligations for stand-alone high-risk systems listed in Annex III — the category that includes hiring, credit scoring and critical infrastructure — were deferred to December 2027 by the Digital Omnibus on AI, Regulation (EU) 2026/1744. AI embedded in regulated products listed in Annex I, such as medical devices, moves to 2 August 2028. Harmonised standards that would tell a developer what conformity looks like in practice are still being drafted. Until they exist, compliance relies on a company's own interpretation of a legal text, which is an expensive position for a small vendor and a comfortable one for a large one with a legal department.

The paper does not claim the strategy has worked. It says intellectual leadership is established and sustained leadership is unproven, dependent on finishing the standards and building enforcement. That is a narrower conclusion than the headlines around it suggest.

For European readers, the practical weight sits in a few places. If you buy an AI service from a non-EU vendor, the vendor's obligations follow the service into Europe, not the vendor's headquarters. If you publish synthetic images or deepfakes, Article 50(4) puts a disclosure duty on you; if you build a generative system, Article 50(2) requires machine-readable marking of outputs. If you build AI tools, the document set you need is larger than it was two years ago, and the deadlines that matter most are December 2027 for stand-alone Annex III systems and 2 August 2028 for AI embedded in Annex I products.

Does the AI Act apply to a company based outside the EU?

Yes, when the system's output is used in the EU or the system affects people in the EU. A US or Asian provider selling into the single market has to meet the same transparency and documentation duties as a European one.

I use an image generator for client work. Do I have to label the results?

If you publish synthetic images or deepfakes, yes: Article 50(4) puts a disclosure duty on you as a deployer. The provider of the generator has a separate Article 50(2) duty to mark outputs in a machine-readable format. Check what your tool embeds, but your own disclosure duty does not disappear if metadata is missing.

When do the strict rules for high-risk AI systems start?

Stand-alone high-risk systems listed in Annex III were deferred to December 2027; AI embedded in regulated products listed in Annex I moves to 2 August 2028. The general transparency obligations already apply, and the Commission's AI Office has been enforcing general-purpose AI documentation duties since August 2026.

The study's own limit is worth stating. It offers conditions for judging whether EU leadership holds internationally; it does not measure how many foreign regulators have actually adopted EU-style rules, and the harmonised standards it describes as necessary are not finished. The December 2027 and 2 August 2028 dates are the next points at which the claim can be tested against something other than a legal text.

Discussion

No comments yet — be the first to share your thoughts.
X

Don't miss out!

Subscribe for the latest news and updates.