OpenAI: evaluation vs approval distinction
OpenAI proposes mandatory CAISI pre-release evaluations for most capable frontier models but opposes CAISI authority to approve or block deployments. Developers retain release decisions.
CAISI should conduct evaluations, recommend mitigations, develop standards, certify independent assessors, coordinate with national security agencies and international partners.
Chris Lehane praised EO direction but pushes greater CAISI role vs NSA-led benchmarking. Mandatory evaluations needed — no lab should decide unilaterally.