Sandbox 03

Bias Detective

Sixteen equally-qualified candidates, split into two groups. Toggle which features a ranking model can see and watch how removing one biased feature isn't enough when a second feature quietly encodes the same bias.

See this sandbox in its lesson context

Which features should the model use?

Avg. rank — Group 1

4.5

Avg. rank — Group 2

12.5

Ranking matches true merit

63%

With both School Tier and Zip Code Tier active, the model is ranking by historical privilege as much as by actual qualifications — even though neither group is more qualified in this data.

RankCandidateGroupGPAExp.SchoolZip
1Candidate DGroup 12.8933
2Candidate AGroup 13.8433
3Candidate HGroup 13.6333
4Candidate BGroup 13.1723
5Candidate GGroup 12.6823
6Candidate FGroup 13.3532
7Candidate CGroup 13.5232
8Candidate EGroup 13.9123
9Candidate JGroup 23.2812
10Candidate NGroup 23.4521
11Candidate LGroup 22.9911
12Candidate PGroup 23.5412
13Candidate KGroup 23.7221
14Candidate OGroup 22.7711
15Candidate MGroup 23.8112
16Candidate IGroup 23.9311