Slowing AI only works if everyone participates
AI is powerful, promising, and dangerous. Slowing down may be sensible, but unilateral restraint shifts power toward less accountable actors and may make the race more dangerous.
More than 1,100 employees from organisations including Anthropic, OpenAI, Google DeepMind, and Meta have signed the Pacing the Frontier statement. Their concern deserves serious attention. AI can greatly advance science, healthcare, education, and productivity. That same power can accelerate cyberattacks, concentrate power, and create systems we do not adequately understand or control.
Not a simple call for a pause
The statement is often presented as a demand to stop AI. That is not quite accurate. Its signatories ask the US government to support an international effort to develop technical and governance tools that could deliberately pace frontier AI progress if this becomes necessary.
That is more sensible than blindly continuing. Yet the hardest problem sits inside the word international.
Unilateral restraint increases risk
A brake only works when every major actor participates in a verifiable way. If American and European companies slow down while labs in China or elsewhere continue, development does not stop. Leadership shifts toward actors subject to less transparency, oversight, and democratic control. The most powerful and least regulated models could therefore become more important and more widely used.
That would produce the opposite of the intended result. Slowing only the relatively more accountable actors does not automatically make the world safer.
The West does not deserve blind trust either
This is not a simple story of a good West facing a bad China. China is an autocratic state, but governments and technology companies in the United States also concentrate enormous power. Political pressure on independent institutions and close ties between government, defence, and AI companies give us reasons to remain critical there too.
Safety must therefore not depend on trust in one country, administration, or company. Rules should be verifiable, independently supervised, and apply to everyone developing models above agreed risk thresholds.
Resilience alongside regulation
AI will not disappear. Alongside international agreements, we need infrastructure capable of withstanding much more powerful AI. This includes stronger digital authentication, information provenance, independent model evaluations, mandatory incident reporting, robust cybersecurity, human fallback procedures, and systems that keep functioning when AI is misused or unavailable.
The companies building this technology hold knowledge that is essential to that work. They should not only ask for a brake. They should actively help make public and digital infrastructure structurally resilient to AI.
Our position
We understand the desire to buy time. Slowing down may become necessary. But without a verifiable global mechanism, a regional pause is not a solution. It cedes ground to the least transparent actors and may make the AI race more dangerous rather than safer.
We do not have an easy answer. Nobody does. That is precisely why safety, international coordination, and societal resilience must now advance as seriously as the models themselves.
Sources: Trending Topics, Pacing the Frontier, Anthropic Frontier Safety Roadmap, and the Bank of England on frontier AI and resilience.
All hands on deck?
Tell us what you're building. We'll tell you what we can ship by Friday.