FairStress tests if AI models drop the correct candidate

Researchers released FairStress, a benchmark (a standard test) for fairness in AI models. It has 300 decision scenarios in 29 areas, such as hiring and lending, and it grows to more than 13 million test cases. The team found that models drop a correct decision more easily when an argument favors a candidate with a minority identity.

How it works

Each scenario gives a model a pair of candidates. The model must select one. The team adds a short persuasive statement. For example, it says that a leader prefers one candidate. Then the team records the selection of the model.

The numbers

With no identity information, models selected the correct candidate at almost the same rate. When identity was more clear, the behavior of models was less equal. The team calls this anti-stereotypical over-correction. A subset of 50K items, FairStress-Core, makes tests take less time.

What is not known

FairStress uses one definition of fairness: equal treatment. An equity view can count the history of discrimination and give other answers. The text names the models but gives no score for each model.

Insight: Fairness tests now look at pressure, not only at answers

The researchers of TypedBench say that a model must have tests for policy, wording and decision results. FairStress shows that equal answers can hide unequal resistance to a challenge. These tests look at the acts of a model, not only its answers. One fairness score from a vendor is weak proof. Buyers who tell vendors to show challenge results get more control.