AI safety is moving toward a discipline built around repeatable measurements rather than broad claims about whether a model is safe or unsafe.
Model evaluations can test capabilities, failure modes and resistance to misuse under defined conditions. But the field still faces difficult questions: benchmarks can become outdated, models may behave differently when connected to tools, and a score measured in a laboratory may not predict performance in a live organization.
NIST’s expanding work on evaluations and AI risk management reflects the need for common methods. Shared terminology, test protocols and reporting practices can help buyers compare systems and give regulators a stronger evidence base for future rules.
The most useful safety programs will combine technical tests with operational controls. Access permissions, logging, human review and incident response remain necessary even when a model performs well on pre-deployment benchmarks. Measurement is not a substitute for governance, but it can make governance more precise.
Source: NIST AI Update