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

Research into AI regulatory frameworks indicates that regulators cannot simultaneously guarantee a fixed risk ceiling and a development deadline when safety research productivity is uncertain, unless low-productivity outcomes are ruled out. A calibrated model demonstrates that maintaining risk ceilings demands a minimum threshold of safety knowledge, and regulations that mandate fixed compute allocations fix delivery timelines rather than mitigating actual risk.

Why it matters

This analysis demonstrates the mathematical incompatibility of simultaneously enforcing hard safety ceilings and rigid deployment timelines under uncertain technical progress. It matters because it reveals that standard regulatory interventions based on fixed compute thresholds fail to control risk, shifting systemic costs onto operational and downstream services.

Key insights

• Simultaneously promising an artificial intelligence risk ceiling and a development deadline is only possible if weak safety research productivity can be ruled out.
• Maintaining a risk ceiling below an unprotected hazard level necessitates establishing a minimum stock of safety knowledge.
• A phased approach of learning first and replaying development with spare compute dedicated to safety requires only 3.3 percent more productivity than the theoretical necessary bound under calibrated baseline conditions.
• Regulatory frameworks that dictate fixed compute allocations control project timelines rather than capping operational risk.
• The financial and operational costs associated with guaranteeing risk bounds under fixed deadlines are ultimately absorbed by customer services.

Source

arXiv — Computers and Society — https://arxiv.org/abs/2610.11093