Regulating AI risks under unknown safety productivity and fixed deadlines

Risk Ceilings and Development Deadlines: Pacing AI under Uncertain Safety Productivity

Computers and Society

Summary

The paper looks at whether it's possible for regulators to promise both a maximum limit on AI risk and a fixed deadline for AI development when they don’t fully understand how productive AI safety research will be. The authors find that to keep risks below a certain level, enough safety knowledge must be accumulated beforehand, which means weak or slow safety research can’t be allowed. They suggest a way to balance learning safety knowledge first and then speeding up development by using leftover computing resources for safety work. Their model shows this approach nearly meets the tightest productivity requirements. They also show that simply setting fixed computing resources earlier mostly controls development dates but not safety risks.

What this means in practice

  • For ai policy makers: Set realistic regulatory timelines that account for uncertain productivity in AI safety research to ensure risk limits are respected.
  • For ai development teams: Adjust compute resource usage to balance safety research and development speed for safer AI deployment under fixed deadlines.

Authors

Li Gan

Abstract

Can a regulator promise both a risk ceiling and a development deadline when safety productivity is unknown? A ceiling below the final model's unprotected hazard requires a minimum stock of safety knowledge, so both promises hold only if weak research can be ruled out. Learning first and then replaying development, with spare compute in safety, comes close to that minimum. In a calibration with only state risk, constant safety yield, and full knowledge transfer, it needs only 3.3 percent more productivity than the necessary bound. Customer services pay for the guarantee. Rules that fix compute allocation fix dates, not risk.
View PDFOpen arXiv