Conceptexternal validity
External Validity
A measure of whether evaluation results can be generalized to real deployment scenarios; in AI safety evaluation it reflects how closely the test environment approximates the real environment.
Timeline (last 90 days)
Oct 23
External validity refers to whether evaluation results can be generalized to real deployment scenarios—if the test environment is too simplified, the model's risky behaviors in the real world may not be triggered; but the closer the environment is to reality, the greater the surface for loss of control brought by tool use and agentic behavior.
Unverified50%