Authored by Bruno Liebhaberg, CERRE Executive Chairman.
When the companies competing to build the most powerful artificial intelligence call for a slowdown, policymakers should listen. But that does not mean European leaders should blithely follow their call for a blanket slowdown. Not being a leader in frontier AI, Europe’s interest is in ensuring neck-to-neck competition between the market leaders. That means it needs a smarter approach: one which combines safety and competition with effective liability for harm.
This month, Anthropic’s Dario Amodei called for “pacing the frontier”, attracting support from OpenAI’s Sam Altman, X’s Elon Musk and Google DeepMind Chairman Demis Hassabis. The appeal followed disclosures that agents powered by OpenAI models had compromised AI platform Hugging Face’s infrastructure during security testing. Former OpenAI and Anthropic researcher Jacob Coxon has also warned of the existential risks of the current pace of AI development.
The motives behind the calls for restraint deserve scrutiny. Perhaps they reflect attempts to maintain investor interest by pointing to models’ immense power and capabilities, following the mantra that ‘all publicity is good publicity’. Perhaps they reflect concern about the sustainability of spending on computing infrastructure and research, or about whether today’s models can deliver the revenues investors expect. Perhaps developers increasingly doubt that existing architectures can be made sufficiently safe and want to buy time to pivot to different forms of AI.
Another possibility is concern about the cost of future compensation claims. AI companies might seek to distinguish risks inherent in use of the technology from harm caused by inadequately tested systems – and try to limit their liability to the latter. But describing a danger as inherent does not settle who should bear its costs.
Then there is the argument that regulation would raise barriers to entry and protect incumbents. That is plausible, but the competitive effects would not be uniform. A slowdown focused on the largest models could leave room for smaller, more efficient and more open alternatives. Primarily, it risks constraining established technology groups seeking to overtake today’s leaders by drawing on their existing services and customer bases.
A global slowdown would also require cooperation from China. Without it, the US might seek to preserve its lead through tighter export controls and pressure on allies to exclude Chinese open-weight or open-source models. That could protect US developers, giving them more pricing power, while reducing Europe’s choice of suppliers. It is a scenario European policymakers should examine, not an outcome they should assume is inevitable.
Europe’s interests lie in affordable access to AI, rapid adoption and the development of world-leading applications. They also lie in avoiding dependence on any one country. Sustained competition between US and Chinese developers helps turn AI models into cheaper, more widely available inputs for European innovation. A safety framework should preserve that possibility rather than unnecessarily restrict it.
What does that mean in practice?
First, Amodei explicitly distinguishes pacing from halting model training. Policymakers should be equally precise: limits on computing power, restrictions on training and longer intervals between releases are different interventions. Each needs a justification based on the risks it addresses, not simply the desire to slow progress.
That does not mean dismissing the dangers. A low-probability loss of control is not reassuring when the consequences could be catastrophic. The aim should be to preserve AI’s benefits while preventing unacceptable harm.
Models with powerful cyber or biological capabilities should therefore require approval before deployment. Models that can automate further AI research warrant particular scrutiny, and less frequent releases of the most advanced models deserves consideration. Independent monitors should have meaningful access to AI laboratories.
Europe already has a basis for oversight in the AI Act. It imposes transparency duties on providers of general-purpose models, with additional obligations on providers of models posing systemic risks to assess and mitigate those risks, report serious incidents and protect cybersecurity. However, the law seems to have gaps: for example, it does not regulate some pre-market research and development, while recent incidents show that such activities can still cause major cybersecurity incidents. The EU should therefore build on the AI Act’s framework.
Second, liability rules should make AI firms bear the costs of harm for which they are responsible, giving them a stronger incentive to test and control their systems. The EU should therefore return to the question of liability. The AI Act is not a comprehensive system for compensating victims. Nor, however, is there a complete legal vacuum: the revised Product Liability Directive expressly covers software, including AI systems, and must be transposed into national law by 9 December 2026.
The Commission withdrew its separate AI Liability Directive proposal in October 2025. That should not end the debate. One could imagine adopting harmonised EU liability rules applicable to frontier models only. This would make responsibility clear when something goes wrong and give people harmed by AI an effective route to compensation.
Compensation after the event cannot substitute for preventing a disaster – so liability rules and more oversight are both needed.
Europe should not have to choose between benefiting from AI and holding developers accountable. Its task is to preserve competition, enable innovation and ensure that those responsible for harm cannot leave others to pay the bill.
