Who regulated AI first?
The claim is true. It is just more interesting than it looks, and for different reasons than the ones usually given.
The missing qualifier
The European Parliament describes the AI Act as the world's first comprehensive AI law, and that stands. The precise version is that the EU was first with a horizontal AI law: one statute cutting across every sector, built on a risk ladder from prohibited to unregulated, applying the same logic to hospitals and to marketing agencies.
Drop the word horizontal and the claim becomes false. China regulated AI years earlier. The difference is that China did it vertically — one regulation for recommendation algorithms, one for synthetic media, one for generative AI — rather than one law covering everything.
That is a real distinction, not a quibble. But it deserves saying out loud, because the detail it conceals is awkward for the story Europe tells about itself.
Who was first at what
Scroll sideways on a small screen.
The detail in the table
Look at the two labelling rows. China's rules on marking AI-generated content took effect on 1 September 2025. The EU's Article 50 began to apply on 2 August 2026.
On the single thing the public actually notices — that AI-made content should say so — the EU came second, by eleven months.
That does not make the AI Act less significant. But it is worth knowing when someone says the EU set the global standard for AI transparency, and it is a precision that almost no English-language explainer includes.
"The EU was the first in the world to regulate AI."
The real distinctionFirst with a horizontal, cross-sector law. China had binding AI rules in force from March 2022, more than two years earlier.
"The rest of the world is following the EU."
The real distinctionOf the five major EU-style laws, one is in force, one collapsed in court, one is dead and one has not moved in two years.
"The AI Act is finished and applies now."
The real distinctionThe EU deferred its own core in July 2026. The high-risk rules do not bite until December 2027 and August 2028.
What happened to the followers
Around 2023 and 2024 the expectation was that the EU model would propagate. Five jurisdictions were heading somewhere similar. Here is where they stand in August 2026.
South Korea is the only one that actually landed, and its law is far lighter than the EU's. A penalty ceiling around 21,000 dollars should be read against the EU's 35 million euros or seven per cent of global turnover. These are not the same kind of regulation; they are the same genre.
Colorado, in full
This one is worth the detail, because it is the clearest single demonstration that the EU model does not spread on its own.
In May 2024 Colorado passed SB 24-205, described at the time as the first comprehensive AI law in the United States. It was visibly modelled on the EU approach: a duty of reasonable care against algorithmic discrimination in consequential decisions, impact assessments, and notification to the state attorney general. It was due to take effect on 30 June 2026.
It never did.
11 December 2025 — the executive order
The White House signs an order directing federal agencies against state AI laws. It establishes an AI Litigation Task Force in the Department of Justice, due to start on 10 January 2026, to challenge state statutes. It ties 42 billion dollars in broadband funding to states not enforcing designated AI rules. Colorado's act is named, alongside two California statutes and a New York City ordinance.
27 April 2026 — the injunction
A federal magistrate blocks enforcement of SB 205 following a constitutional challenge brought by xAI. The Department of Justice intervenes on the company's side.
14 May 2026 — repeal and replace
Two weeks later, Colorado signs SB 26-189, repealing its own law and replacing it. The impact assessments are gone. The substantive anti-discrimination framework is gone. What survives is largely a notification duty for consequential decisions.
1 January 2027 — the replacement takes effect
A materially weaker statute than the one the state passed in 2024, and one that no longer resembles the EU model it was drawn from.
Two years from first-in-the-nation to repealed, without a day of enforcement. Whatever one makes of the merits, it is not the trajectory the Brussels-effect thesis predicts.
The Brussels effect: two numbers pointing opposite ways
The Brussels-effect thesis holds that EU rules go global because it costs large firms more to build two versions of a product than to apply the strictest rule everywhere. It happened with the GDPR. The question is whether it happens with AI.
For. A review of more than 100,000 data points from nearly 3,000 companies found that 47 per cent of the firms referencing the AI Act in their reporting are headquartered outside the EU. American companies alone account for 35 per cent of non-EU citations, from a country with no federal AI statute. British, Swiss and Norwegian firms make up roughly a quarter. Microsoft, Google, OpenAI and xAI have all voluntarily signed parts of the EU's codes of practice to keep market access.
Against. The same review found that only 12.4 per cent of companies globally require human review of individual AI decisions, and that fewer than one in four assess the impact on workers' rights. Naming the AI Act in an annual report is risk language. It is not the same as changing how the work is done.
There is a theoretical objection too. Ben Crum at Vrije Universiteit Amsterdam has argued that AI lacks the network effects that made the GDPR a global standard: AI applications spread through a far more fragmented market, which makes it easier to build a separate European version than it ever was to run two privacy regimes.
And then the least comfortable objection: that the EU has retreated from its own position. The amending regulation that entered into force on 27 July 2026 deferred the high-risk rules by between eighteen months and two years, softened the AI literacy duty, and simplified obligations for smaller firms. Max Schrems called the wider package the biggest attack on Europeans' digital rights in years. Others point to the Draghi report and Europe's lag in technology sectors and argue the correction was overdue. Both readings have support.
The honest summary
The EU was first with a horizontal AI law, and that is a genuine achievement. It remains the most ambitious attempt anyone has made to regulate AI across an entire economy, and it is the reference point every other regime is measured against — including the ones that reject it.
But being first is not the same as being followed. China came earlier on labelling. South Korea chose a far lighter model. Colorado abandoned the EU approach explicitly, after being sued. Canada dropped its bill. Brazil has not moved in two years. And the EU deferred its own core before it took effect.
The effect is therefore strongest as agenda — everyone argues in terms of risk tiers and transparency duties because the EU wrote the vocabulary — and weaker as substance. That is a more careful conclusion than either "the Brussels effect is a myth" or "the rest of the world is catching up", and it is the only one the numbers support.
What this means if you are building something
Two practical consequences.
First, you can no longer assume AI regulation converges. Selling into both the EU and the US means two sets of requirements that differ not only in strictness but in structure. Planning for one and hoping it covers the other is how you end up compliant in neither.
Second, the European direction is still the strictest, and therefore still the sensible one to build against if you are only going to pick one. Not because everyone else will follow, but because whoever satisfies the EU rules rarely breaches anyone else's.
For what actually applies to you today, start with what the AI Act requires now.
Sources
- European Parliament on the AI Act — 'the world's first comprehensive AI law'
- White & Case AI regulatory tracker: China — dates for the Chinese regulations
- Library of Congress on China's deep synthesis rules — in force 10 January 2023
- Loeb & Loeb on China's labelling measures — in force 1 September 2025
- Cooley on South Korea's AI Basic Act — entry into force, grace period and penalties
- Library of Congress on South Korea — independent confirmation
- McDermott on Colorado — the injunction and the replacement act
- Baker Botts on the Colorado repeal — what SB 26-189 changed
- Department of Justice on intervening in the xAI suit — the department's own announcement
- Paul Hastings on the executive order — the litigation task force and broadband funding
- StateScoop on the state moratorium — the moratorium left out of the defence bill
- Deep Lex: AI regulation in Brazil — status of PL 2338
- Montreal AI Ethics Institute on AIDA — what happened to the Canadian bill
- Council of Europe Framework Convention on AI — opened for signature 5 September 2024
- Euronews on the Brussels effect — the 47 per cent and 12.4 per cent figures
- Internet Policy Review: Brussels effect or experimentalism? — the theoretical objection
- Lawfare on the EU's change of course — the omnibus, the Draghi report and the criticism
- Regulation (EU) 2026/1744 — the amending regulation, in force 27 July 2026