How to check.
WHAT TO MEASURE
Outcomes by group, not overall accuracy.
Approval rates, error rates, false positives and false negatives, each broken down.
WHAT TO LOOK FOR
Substantially different outcomes between groups with similar underlying characteristics.
WHY OVERALL ACCURACY MISLEADS
A system can be highly accurate overall while performing poorly for a minority group.
The aggregate hides it.
WHAT TO DO WITH DIFFERENCES
Investigate before accepting. Some differences reflect genuine underlying differences; many do not.
The burden is on establishing which.
WHAT IF YOU CANNOT TEST
If you cannot measure outcomes by group, you cannot claim the system is fair.
That is a reason for caution, not a reason to proceed.
FOR SMALL ORGANISATIONS
You may lack the data to test meaningfully.
That argues for human decision-making rather than automated.
WHAT TO DOCUMENT
The testing performed and what it showed.
If a decision is challenged, that record matters.
