AI Safety Requires More Than Slowing the Pace of Progress

Summary

In this opinion essay, computer scientist Stuart Russell examines Anthropic CEO Dario Amodei’s proposal to “pace the frontier” of AI development. Amodei’s 3,800-word letter responds to concerns about AI advancing rapidly through recursive self-improvement and has received support from Sam Altman, Elon Musk, Demis Hassabis and Satya Nadella. The proposal would place independent third-party evaluators inside AI companies with full system access; encourage frontier companies in democratic countries to adopt common safety standards and limits on unchecked progress, backed by regulation where necessary; and extend the arrangement to authoritarian countries through a broader compact. Amodei says pacing would not stop training or technical progress and could create one to two years for work on interpretability, alignment and testing. Russell rejects the idea that a slower timetable alone can make development safe. He argues that developers should first define non-negotiable safety requirements and continue only when models demonstrate the required alignment properties, an approach he identifies with researchers’ “red lines” proposals. If those requirements cannot be met, progress should stop altogether, equivalent to a red flag rather than a Formula 1 pace car. Russell links recursive self-improvement to the possibility of an irreversible loss of human control and says the acceptable risk level should be far below the levels he attributes to current AI executives. He also points to the lack of credible plans for controlling superintelligent systems, warns that past investment does not justify continuing an unsafe path, and says current attention to AI risk and upcoming political discussions create an opportunity to choose a different course.