Federal AI testing agreements would require development disclosures and capability tests

The federal standards agency would have to include disclosure requirements and capability tests in covered AI testing agreements. The director would decide which AI systems count as frontier systems covered by these rules. Developers would disclose information upon the director’s request about AI use in developing the system. That information could include business metrics, estimates of work done without human review, and methods used to make those estimates. The requirements would apply when the director enters an agreement with a developer to evaluate a covered system before release.

The National Institute of Standards and Technology would assess trends in AI capabilities. The assessment would cover capabilities that might contribute to scientific research breakthroughs. The agency would use information disclosed under covered testing agreements, if any, along with other information.