What is a measurable safety standard?

Definition

A measurable safety standard defines observable requirements, test conditions, thresholds, and evidence for judging whether a system meets a safety goal. It replaces a vague statement such as being safe with criteria that evaluators can apply consistently.

The measurement must cover relevant failure modes and operating conditions. A convenient metric can create false confidence if it is easy to pass but does not represent the real harm or behavior the standard was meant to address.

ELI5

A measurable safety standard is a clear rule for checking safety instead of simply saying that something should be safe. It explains what to test and what result is acceptable.

For example, a standard for a medical AI tool might require it to flag uncertain cases for qualified review and measure how often it fails to do so on an independent test set.

What should a measurable safety standard include?

It should include the intended risk, test method, operating conditions, evidence, threshold, and process for handling failures.

Can one safety metric cover every risk?

Usually not. Different risks require different tests, and combined evidence is often more informative than one headline score.

  1. Nick Saraev, Jack Roberts and Jakub Pachocki with AI safety and compute infrastructure themes
Definition card for measurable safety standard: What is a measurable safety standard?

A measurable safety standard turns a general AI safety goal into specific criteria that can be tested and compared.