Any AI that affects people needs checking for consistent treatment. This records the tests, the results, and the actions, so fairness is evidenced rather than assumed.
What's Included
- Tests defined before you run them so results are not chosen afterwards.
- Outcomes compared across groups on the same measure.
- Findings and actions recorded including what you decided to accept.
- An assistant that highlights differences worth investigating.
How To Use
- Define the tests in advance. Choosing measures afterwards proves nothing.
- Compare outcomes, not intentions.
- Record what you accepted and why, not only what you fixed.
- Involve someone qualified. This records the work, it does not perform it.
Trigger it from risk assessment, and record actions in the incident register.