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 contextWhich 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.
| Rank | Candidate | Group | GPA | Exp. | School | Zip |
|---|---|---|---|---|---|---|
| 1 | Candidate D | Group 1 | 2.8 | 9 | 3 | 3 |
| 2 | Candidate A | Group 1 | 3.8 | 4 | 3 | 3 |
| 3 | Candidate H | Group 1 | 3.6 | 3 | 3 | 3 |
| 4 | Candidate B | Group 1 | 3.1 | 7 | 2 | 3 |
| 5 | Candidate G | Group 1 | 2.6 | 8 | 2 | 3 |
| 6 | Candidate F | Group 1 | 3.3 | 5 | 3 | 2 |
| 7 | Candidate C | Group 1 | 3.5 | 2 | 3 | 2 |
| 8 | Candidate E | Group 1 | 3.9 | 1 | 2 | 3 |
| 9 | Candidate J | Group 2 | 3.2 | 8 | 1 | 2 |
| 10 | Candidate N | Group 2 | 3.4 | 5 | 2 | 1 |
| 11 | Candidate L | Group 2 | 2.9 | 9 | 1 | 1 |
| 12 | Candidate P | Group 2 | 3.5 | 4 | 1 | 2 |
| 13 | Candidate K | Group 2 | 3.7 | 2 | 2 | 1 |
| 14 | Candidate O | Group 2 | 2.7 | 7 | 1 | 1 |
| 15 | Candidate M | Group 2 | 3.8 | 1 | 1 | 2 |
| 16 | Candidate I | Group 2 | 3.9 | 3 | 1 | 1 |