Each project arrives with its purpose, milestones and first tasks written out. Here is how the first three read.
Inventory of models and automated decisions
This is a team plan, best installed into a shared workspace so everyone with a role in it can see and own their tasks. Most organisations cannot say how many models make or shape decisions about people, and bias reviews then only reach the systems someone happened to mention. A single inventory listing each model, the decision it feeds, who is affected and who owns it is the scope every later review depends on.
First tasks
- List the models and scoring rules your own team runs today
- Ask procurement for every vendor tool that scores or ranks people
- Write three criteria that make a decision high risk
- Tier each entry and name its business owner
- Add newly proposed models and vendor tools to the inventory @recurring(monthly:9)
Register of groups and sensitive attributes to test
Every bias review needs to know which groups it is comparing, and the answer depends on the decision, the jurisdiction and the population affected. Writing down the characteristics in scope, such as sex, age, ethnicity, disability and their proxies, with counsel's view on which are legally protected where you operate, stops each analyst choosing their own list.
First tasks
- Draft the characteristics you think each high-risk decision should be tested on
- Ask counsel which are legally protected for these decisions and places
- Note likely proxies for each characteristic in your data
- Record where each attribute comes from and who may see it
- Ask counsel whether the protected list has changed @recurring(yearly)
Fairness metric choice for each decision type
Demographic parity, equal opportunity, predictive parity and calibration cannot all hold at once when base rates differ, so choosing a metric is a value judgement, not a technical default. Recording which metric applies to each decision, and why, means a disparity finding can be argued about on its merits rather than on the yardstick.
First tasks
- Class each high-risk decision as benefit, burden or ranking
- Write down which error hurts people most for each class
- Choose a primary and secondary metric and justify each in a paragraph
- Agree a tolerance for each metric with the decision owner