Kapoor and Narayanan argue against extraordinary AI rules

Sayash Kapoor and Arvind Narayanan published a long argument on May 21 against what they call "extraordinary government interventions" on AI: policies that bypass normal lawmaking, restrict actors who have not caused harm, and act on precaution rather than evidence. On their AI as Normal Technology newsletter, the two contrast that approach with one they prefer: building societal resilience to AI risks through ordinary policymaking, the way cybersecurity threats were eventually managed.

Their core case is that nuclear nonproliferation, often the analogy reached for when people argue for AI export controls and compute caps, depended on physical bottlenecks AI does not have. Adversaries can match a frontier model's capability within months once it ships, so a control regime tilts into "constant erosion" rather than a steady state. The cybersecurity comparison is the constructive half: worms cost the economy billions in the 2000s, but the response was bug bounties, vulnerability disclosure, and automated testing, not turning off parts of the internet. They argue the same posture, applied to AI, would cover the actual risk surface without the political cost of restricting access.

The piece arrives in a moment when bipartisan items like transparency reporting and safety-research protections still sit unenacted, the gap Kapoor and Narayanan want filled by ordinary work rather than bypassed. Read the full argument.

Why it matters

If you work on AI policy, model release, or red-teaming, the resilience framing changes what counts as success: not blocking releases, but funding the defensive stack (eval tooling, vulnerability reporting, transparency rules) that lets ordinary law catch up. For builders, it predicts which restrictions are likely to stick and which will erode.