The Lawfare Podcast · The Lawfare Institute

Scaling Laws: Can AI Make AI Regulation Cheaper?, with Cullen O'Keefe and Kevin Frazier

March 6, 2026·53 min·5 clips

Compute thresholds for AI regulation are failing as new models trade training compute for cheaper inference.

The Lawfare Podcast presents an episode from its new series Scaling Laws, hosted by Alan Rosenstein of the University of Minnesota. The guests are Cullen O'Keefe, Research Director at the Institute for Law and AI, and Kevin Frazier, AI Innovation and Law Fellow at the University of Texas School of Law. The episode centers on a new paper by O'Keefe and Frazier arguing that AI itself can dramatically lower the costs of complying with AI regulation, a concept they call automated compliance. Frazier opens by framing the compliance cost problem by company size. Large firms like Meta and Google have compliance teams numbering in the hundreds or thousands, making regulatory costs a de minimis fraction of revenue. Startups, by contrast, typically cannot afford to hire attorneys as early employees. A single privacy policy update via outside counsel can cost $5,000, a significant amount against average startup operating expenditures of $55,000 per month. This means many startups simply ignore new regulations until they face enforcement. O'Keefe explains how the AI safety community has traditionally addressed this disparity through compute thresholds, where regulations only apply to models trained above a certain number of floating-point operations. This approach targeted well-capitalized firms while exempting smaller players. However, the emergence of reasoning models has undermined this approach: these models trade training compute for inference compute, meaning a cheaper-to-train model can match the capabilities of a more expensive one by thinking longer at query time. Combined with falling compute costs and ongoing efficiency gains, fixed FLOP thresholds are becoming increasingly disconnected from both capability levels and the financial capacity of the firms they are meant to target. Frazier adds that compute thresholds also miss emerging risks like AI companions that can drive harmful user behavior without requiring large training runs. The core proposal is automated compliance: using AI systems to handle regulatory tasks including compiling transparency reports, performing automated evaluations of AI models, monitoring safety and security incidents, and filing incident disclosures with regulators. The hosts discuss how this echoes Paul Ohm's recent paper "Toward Compliance Zero" and represents a natural consequence of AI expanding the production possibility frontier. The discussion addresses several objections. On Goodhart's Law concerns, Frazier argues that automated compliance does not change a company's operations or create reward incentives for gaming metrics; it simply automates background reporting tasks. On whether the real objection from industry is substantive rather than cost-based, O'Keefe argues that lowering compliance costs allows the policy debate to focus on first-order questions about what safety levels society actually wants. The most novel proposal is what they call "automatability triggers," a form of sunrise clause where a regulation is enacted but does not take effect until AI-powered compliance tools can implement it below a cost threshold set by the legislature. The trigger would be certified by a body like the Secretary of Commerce or NIST. This mechanism addresses a tension between those who want regulation now (to prevent catastrophe) and those who worry about premature path-dependent regulation locking in incumbents or steering technology development in suboptimal directions. Frazier notes that sunrise clauses are not novel in law, citing examples in privacy regulation and occupational licensing. The episode also addresses who would build these compliance tools, with O'Keefe noting that antitrust law would prevent coordinated refusal, and open-source models provide a backstop against intentional holdbacks by frontier labs.

Listen to the show on

Apple Podcasts Spotify

Play clips from this episode More from The Lawfare Podcast