At Eumans, we have spent the past two years arguing that how artificial intelligence is built cannot be separated from who gets to decide. AI can strengthen public services, widen access to knowledge, and help citizens take part in the decisions that shape their lives. But we have also argued that a technology this consequential should not be developed on terms set almost entirely by a handful of companies and two governments.
This month, that argument stopped being abstract. Some of the people building the most advanced AI systems in the world said publicly that development is moving faster than our ability to keep it safe. We take that warning seriously. We also think it reveals something important about where the danger comes from. Europe now has both a reason and an opportunity to build AI differently. Two events have convinced us of this.
What Happened
The first was a resignation. In the second week of September, Jacob Coxon, a 27-year-old British researcher who had spent roughly three years on pretraining work at OpenAI and then Anthropic, quit and wrote that neither company was acting responsibly and that both were “gambling with our lives.” What gave his warning weight was its source: he had helped build the systems he now fears. Speaking to Axios, he was careful to say he had not seen Anthropic compromise safety to beat its rivals. His concern was the pressure itself, which in a race pushes teams to skip steps in oversight. Speaking to TIME, he described colleagues who share his worries but have come to accept the race as a given.
The second was a response from the top of the industry. On 12 September, Anthropic’s chief executive, Dario Amodei, published an essay titled “We Must Pace the Frontier“, arguing that “we must slow the pace at which we improve the capabilities of AI models.” He pointed to AI’s growing ability to help build its own successors, and to an incident, reported by Euronews among others, in which OpenAI models under testing broke out of their environment, reached the internet and got into Hugging Face’s systems. He proposed three steps: independent evaluators embedded inside AI companies with employee-like access, shared safety standards among labs in democratic countries, and, where possible, coordination with China. Anthropic committed to the first step immediately. Reuters reported that OpenAI’s Sam Altman and Elon Musk both backed the call, a rare moment of agreement among rivals.
To be clear, we think this is an important moment, and Amodei deserves credit for saying something uncomfortable in plain terms. But the reaction to his essay tells us as much as the essay itself.
Why the Race Sets the Pace
Consider the limits the proposal places on itself. Altman has since stressed, as CNBC reported, that pacing does not mean stopping. Amodei writes that any slowdown among democratic labs cannot go further than America’s lead over China, because slowing more would let Chinese projects pull ahead. His reasoning is that a wider lead creates room to pace safely and more leverage for a future agreement. He also argues that the most effective route is regulation covering every American frontier company, with voluntary coordination in the meantime because laws take time to pass.
Then consider who answered. Reuters reported that President Trump dismissed the warnings as exaggerated, as the United States races to keep its lead over China. The White House science adviser David Sacks told the labs, according to NPR, that they were free to slow down on their own if they thought it necessary. Beijing pushed back too. At a regular briefing, China’s foreign ministry warned that fearmongering and “vicious competition” would only disrupt global AI governance, responding in part to Amodei’s call for the United States to keep restricting China’s access to advanced chips. And when markets opened on Monday, Reuters reported that European technology shares fell around 2%, with ASML, Infineon and Soitec among the hardest hit.
Brussels offered a fourth response, and it split the difference. Rather than dismiss the warnings, von der Leyen leaned into them. Europe would, she said, “step up together with our like-minded partners like Canada, the United Kingdom and others” on model evaluation, verification, early warning and AI security. She also announced she would invite the frontier labs in to discuss how the EU could support their own efforts to pace the frontier.
Yet, for this to be successful, she said Europe needs to “massively boost our computing capacity” so that it can “stay in the race.” That is the tell. Even the Commission’s own language keeps Europe inside the competitive logic Amodei’s essay describes, racing to catch up on the same frontier-model infrastructure, rather than building a genuine alternative to it.
We draw three conclusions from this:
● The pressure is structural. Amodei himself warns that commercial incentives can fuel a race to the bottom that makes these risks more severe, and Coxon’s account from inside the labs points the same way. This is not a story about bad people. It is a story about a system that rewards speed.
● Restraint remains fragile. Until the regulation Amodei calls for exists, the pace of development depends on the judgement of founders and the patience of investors, and the US administration has signalled it does not intend to impose limits.
● Europe is exposed either way. Monday’s selloff showed that European companies are tied to the American race whether or not Europe has any say in its terms.
We want to be fair to the American model. Its competition, its capital and its appetite for risk produced this technology and many of its benefits. The problem is not competition as such. It is that decisions affecting everyone are being made inside a contest where no one at the table is there to represent the public.
Europe’s Choice
Europe is not immune to the pull of that contest. Under the Digital Omnibus on AI, which entered into force in July, the EU pushed back its main rules for high-risk AI systems to December 2027, after implementation fell behind schedule and leading European companies had pressed for a pause. Von der Leyen’s State of the Union speech called the AI Act “a crucial piece” of putting guardrails in place, even as the Commission spent the summer extending the deadline for its toughest rules. The temptation to simply try to keep up is real.
But Europe is not at the very front of the frontier race, and that gives it room to choose. The United States builds AI around the individual as a consumer, and China builds it around the individual as a political subject. Europe should build it around the individual as a citizen: someone who governs, rather than someone who is governed or sold to.
How Civic AI Answers This Moment
Our proposal for a European Civic Artificial Intelligence was not written in response to this month’s events, but several of its elements speak directly to them:
● Openness as a foundation. Amodei proposes embedded evaluators partly because, as he acknowledges, companies still decide what the public gets to see. We start from the other end: an AI ecosystem that is federated, open source and interoperable by default, trained on data the public can inspect, so that no single actor, public or private, can capture the technology that runs democratic life.
● Shared public infrastructure. Today’s race is a contest among the few firms that can afford frontier-scale computing. We propose a “CERN for AI,” with €100 billion in EU investment between 2027 and 2032 to connect existing supercomputers, AI Factories, data spaces and research centres into shared infrastructure open to researchers, civic technologists, small innovators and public administrations. Alongside it, a European Sovereign Fund for AI and Data would direct investment toward cooperatives and public interest ownership rather than concentrated shareholder power.
● Many models, not one. We favour a federated ecosystem of smaller, local models, organised around federation and subsidiarity, so that local systems solve local problems and no single, all-powerful AI replaces the many. We do not claim this alone resolves the risks at the global frontier. It does offer Europe a path to useful, sovereign AI that does not depend on winning a race it did not start.
● Deliberation from the start. Amodei writes that society must have a say in how this technology is used, and that slowing down would allow more time for public deliberation. We agree, and we believe deliberation should shape the rules from the beginning rather than fill the time a company manages to buy. In fact, society should decide what values are written in the very code of these general-purpose technologies, as the digital world is already as concrete as the physical one.
Europe already has the tools to act. The proposed AgoraEU programme, with up to €9 billion foreseen for democracy in the 2028 to 2034 EU budget, is a natural place to start. If Europe is serious enough to invest in AI for defence, it must be equally serious about investing in AI for democracy. Our European Citizens’ Initiative for a Civic AI, which needs one million signatures from at least seven member states, is how citizens can put that choice directly before the European Commission.
The warning from inside the AI industry is right: the race is moving faster than our ability to govern it. Europe’s answer should not be to join that race on worse terms, but to build AI that citizens can see, share and shape.
