A one-to-two-year AI slowdown is not a safety plan
Anthropic's 'pacing the frontier' proposal has drawn high-profile backing, but critics say pausing for a year or two still leaves humanity without a real answer to the question nobody has solved: how do you keep an advanced AI system under control?
Key points
- Jacob Coxon resigned from Anthropic, the AI safety company, in a move that brought AI risk back into mainstream public debate.
- Anthropic CEO Dario Amodei has proposed "pacing the frontier", a plan to slow AI development by one to two years to allow safety research to catch up.
- Both Sam Altman (CEO of OpenAI) and Elon Musk (founder of xAI) have publicly backed the Amodei proposal.
- A short pause does not answer the core problem: nobody currently knows how to guarantee that a powerful AI system will remain under human control.
A resignation letter can sometimes do what years of expert warnings cannot. When Jacob Coxon left his role at Anthropic, the AI safety company behind the Claude family of AI assistants, it pushed AI risk from conference rooms into dinner-table conversations. Millions of people are now asking, for the first time, whether AI companies have been taking risks they shouldn't.
Dario Amodei, Anthropic's CEO, responded with a proposal he calls "pacing the frontier": slow down the most powerful AI systems by one to two years, buying time for technical safety research to mature. Sam Altman and Elon Musk have both endorsed the approach. We covered the broader coalition behind that push on 16 September in "AI bosses are calling for a slowdown. Nobody knows what that means."
That sounds reassuring. It isn't, quite.
What's actually missing from the plan?
A pause buys time only if you know what to do with it. The field does not have a proven method for ensuring that a sufficiently powerful AI system stays aligned with human intentions, meaning it keeps doing what people want rather than pursuing goals of its own. That gap is the problem, and a one-to-two-year slowdown does not close it.
AI safety researcher David Krueger, writing in The Guardian AI, argues that what the world needs is not a shorter runway to the same destination, but a concrete, verifiable plan for confirming that an AI system won't cause catastrophic harm before it is deployed. Amodei's phrase "profound progress" on safety is not the same thing as a solution.
If engineers discovered a flaw in a new aircraft design but weren't sure how to fix it, grounding the fleet for eighteen months would help only if the repair was actually identified in that window. No fix found, same problem waiting.
What would a real plan look like?
Safety researchers generally point toward two directions worth serious investment. One is interpretability research, work that tries to understand what is actually happening inside an AI model's calculations so engineers can spot dangerous behaviour before it surfaces. The other is developing clear, testable standards a model must meet before it gains more capability or a wider release, the equivalent of airworthiness certificates for AI.
Both areas are genuinely hard and underfunded relative to capability research, the work that makes models smarter and faster.
For ordinary people, the practical implication is blunt: the companies building the most powerful AI systems don't yet have a complete answer to the safety question, and the most prominent public proposal on the table buys a little time without solving the underlying problem. That is worth knowing when evaluating any headline about AI breakthroughs.
A short pause might still be better than none. But treating a delay as a solution would be a costly mistake, and on this beat, the difference between those two things is exactly what's being glossed over right now.
