Can AI invent its own stereotypes? In a hiring game, it did.
In a Princeton and University of Chicago simulation, AI models' hiring became far more segregated by group than people's, though every candidate was equally good.
In a Princeton and University of Chicago simulation, AI models' hiring became far more segregated by group than people's, though every candidate was equally good.
· Captions, from MIT Technology Review, Gizmodo, SHRM
Night-city illustration drawn in code by AI Broadsheet. Music: “Lofi HipHop Melody Loop 99 BPM” by holizna (CC0, Freesound).
What the video says
Can AI invent its own stereotypes? In a hiring game, researchers at Princeton and the University of Chicago found it can. AI models hired for 20 jobs in a made-up city, from four made-up groups. Every candidate was equally good; the models weren't told that.
Hiring after hiring, the models' choices became segregated by group. On a segregation scale, humans scored 0.84; the AI models scored about 65% higher, and o3 scored 1.83, near the maximum.
Lead researcher Ryan Liu says the models are eager to generalize from limited data. Telling the models to be fair had limited impact; rewarding diverse hiring made them less biased. It was a simulation with fictional groups, not a test of real hiring tools. Would you trust AI with your résumé?
Tell us below.
Sources
Night-city illustration drawn in code by AI Broadsheet. Music: “Lofi HipHop Melody Loop 99 BPM” by holizna (CC0, Freesound). Report a mistake