The EU wants to slow down the AI that Europe isn't building
When Ursula von der Leyen delivered her State of the Union address in Strasbourg on 16 September, she said something no Commission president has said from that podium before: that the development of the most advanced AI models should be slowed down.
The reason she gave was not competition or jobs. It was security. The models now being built will, in her words, enable hacking at a level no one thought possible — and they will soon be in the hands of adversaries. She also pointed out that the AI labs themselves have told the EU it is time to slow down on self-recursive models and pace the frontier.
A European head of government repeating the frontier labs' own warning in a formal address is a shift in tone more than in law.
What was actually announced
- Cooperation with Canada and the United Kingdom on model evaluation and verification, early warning and AI security.
- An invitation to the leading AI labs to discuss slowing development.
- Sector initiatives in health, transport, agriculture, manufacturing and defence and space, expected in November.
- Enforcement of the AI Act as the central guardrail.
And what was not announced: no deadlines, no enforcement mechanisms, no figures — and no date for the meeting with the labs, which is the very instrument meant to address the risk she warned about.
The problem with setting rules for someone else's technology
The day after the speech, Euronews put numbers on the contradiction. The EU attracts less than 5 percent of the roughly one trillion dollars invested in AI globally. The US accounts for 75-80 percent, China for 10-14. Europe's largest AI company, France's Mistral, is valued at 24 billion dollars — against OpenAI at 852 billion and Anthropic at 965.
Irish MEP Michael McNamara called the pledge lip service: most of the developers the EU would like to slow down operate outside Europe and pay little attention to European rhetoric.
What remains is the EU's real lever: market access. The AI Act applies to anyone who wants to sell in Europe, wherever the model was trained. That is a slender weapon next to owning the technology — but it is not nothing.
Why this matters to you too
The part of the speech about hacking is not distant science fiction. AI-driven attacks already reach ordinary people as convincing phishing, voice cloning and deepfakes. We have collected the most important advice in Hackers are getting more numerous and better.