Scaling Laws Can AI Make AI Regulation Cheaper?, with Cullen O'Keefe and Kevin Frazier
Feb 24, 2026
Kevin Frazier, AI innovation and law fellow who studies AI governance, and Cullen O'Keefe, research director focused on frontier AI policy, discuss automating regulatory compliance. They explore how AI can compile reports, run model evaluations, and detect incidents. They debate limits like Goodhart's Law, the fairness of compute thresholds, and the idea of conditional "automatability triggers" to delay enforcement until cheap tools exist.
AI Snips
Chapters
Transcript
Episode notes
Goodhart Concern Is Manageable For Routine Metrics
- Goodhart's Law is a risk but automated compliance often just collects existing background data rather than creating new gaming incentives.
- Kevin Frazier argues automation replicates routine tasks rather than rewarding firms to change behavior.
Automation Lowers The Administrative Tradeoff
- Automated compliance can reduce administrative friction, letting debates focus on the substantive safety vs innovation trade-offs.
- Cullen O’Keefe frames this as shifting the production-possibility frontier so safety can be achieved with less innovation loss.
Automatability Triggers Tie Enforcement To Cheap Compliance
- Automatability triggers delay enforcement until compliance costs drop below a legislated threshold, creating a conditional sunrise clause.
- Cullen O’Keefe describes legislating that a rule becomes effective only when AI-driven task costs fall to X dollars.