Bias Testing
The process of testing an AI model for accuracy and outcome differences across groups, including protected classes and proxy variables.
Bias testing, also called fairness testing, is the process of checking whether an AI model produces systematically different outcomes across groups. It usually includes accuracy testing, outcome analysis across protected classes, and protected-class proxy screening.
The regulatory sources describe it in their own words rather than in a single term of art. The NAIC Model Bulletin asks for “bias analysis and minimization” in an insurer’s data practices, and for methods that detect and address unfair discrimination resulting from a predictive model; it prescribes no particular test. NYDFS Circular Letter No. 7 sets a cadence instead of a method: test for unfair or unlawful discrimination before the system reaches production, keep testing on a regular schedule, and test again after any material change to the data or the model. For the quantitative side it names six candidate metrics: adverse impact ratio, denials odds ratios, marginal effects, standardized mean differences, z- and t-tests, and drivers of disparity. It adds that insurers are not expected to collect additional data about individuals in order to run them.
A complete bias-testing program records the groups tested, the metrics used, the results, and any actions taken. See our glossary entries on algorithmic bias, disparate impact, and the proxy test.