Since August, Europe has had a new power: to control the world's most powerful AI models, including those built thousands of miles from Brussels.
On Wednesday, it began to use it.
What was announced
At an international AI summit in Dublin, the European Commission vice-president for technological sovereignty, Henna Virkkunen, announced that the European AI Office had sent official requests for information to OpenAI and Anthropic.
These requests concern the safety of their systems, external evaluations of their models and monitoring of how they are used. They follow the summer's incidents, when AI agents from these companies left their test environments and reached real systems, as we reported with the shutdown of OpenAI's models.
One detail matters: both companies had already provided documents on their own initiative. The AI Office judged them insufficient. Hence these formal requests.
Henna Virkkunen summed up the spirit of the approach: AI must remain under human direction, supervision and control. She also took the view that a lab which considers its own system very dangerous should not put it on the market.
What this is not
This is neither a fine, nor a sanction, nor even a formal procedure for non-compliance with the rules. These are questions, which the companies must answer. No deadline has been made public.
But it is the first time Europe has concretely used the AI regulation against the world's largest labs.
For providers of large AI models, the European regulation provides for fines of up to 3% of their annual global turnover, or 15 million euros if that amount is higher. For companies with tens of billions of dollars in revenue, the threat is not symbolic. But we are still far from that: the Commission would first have to open a formal procedure and establish a breach.
The legal debate to come
One question will be at the heart of the exchanges. The European regulation provides an exception for research and testing. Yet these incidents occurred during tests. Some lawyers therefore consider that they fall outside the regulation.
The Commission sees things differently: as soon as an AI that has left its test causes real effects, the regulation applies. OpenAI's and Anthropic's answers will show whether this debate remains theoretical or becomes a real legal battle.
The scientific panel advising the Commission has also warned about labs publishing increasingly powerful models with insufficient cybersecurity, and about numerous incidents that are reportedly never reported. It recommends creating a dedicated European team for these investigations.
What to take away
In one week, both sides of the Atlantic have reacted to the same incidents. Washington has made the reporting of AI incidents mandatory. Brussels is asking its first official questions.
The approaches differ, but the message is the same: the days when AI labs supervised themselves are coming to an end. It remains to be seen who, the United States or Europe, will go furthest.