Risk ceiling and deadline: an economist shows that a regulator cannot promise both without knowing the productivity of safety research

In a formal model, ensuring both a maximum level of risk and a development deadline requires ruling out the assumption of inefficient safety research.

Can a regulator impose both a risk cap and a development deadline on an AI lab, when no one knows how fast safety research is progressing? In an article published on arXiv, economist Li Gan responds with a formal model. If the cap is set below the level of danger that the final model would have without protection, a minimum stock of safety knowledge must be accumulated; both commitments therefore only hold if unproductive research can be excluded.

The strategy that comes closest to this minimum is to learn first, then to replay the development by devoting the available calculation to safety. In a simplified calibration, it requires only 3.3% more productivity than the theoretical limit. The author draws a direct conclusion for public policy: the rules that set the distribution of the calculation between capacity and security set dates, not a level of risk.