Credit, hiring, housing: 18 language models give their worst marks to white men without a college degree

Out of 17,280 mock file evaluations, this group has the lowest average score of the eight profiles tested, while Black women with degrees receive the highest.

Economist Maxim Chupilkin submitted 18 language models, from 12 developers, to experiments based on fictitious files in three situations: a credit application, a job application and a rental application. Only the gender, origin, age, nationality and educational level varied, with the financial or professional situation remaining the same in each scenario. Each model rated the 32 profiles ten times per scenario, representing a total of 17,280 notes.

A group masked by the big categories

On average, white men without a university degree arrive last of the eight groups crossing gender, origin and degree: 75.87 out of 100 for credit, 92.71 for hiring and 86.62 for housing. Graduated black women get the highest scores, with differences of 2.9 to 3.7 points. Taken in isolation, each attribute favours women, black candidates and graduates in all three situations, and the group of unqualified white men is last or penultimate in 46 of 54 model-scenario combinations.

The author draws this result closer to the growing economic and health fragility of this population, and points out that an audit of bias limited to gender or origin categories alone would not have revealed this.