AI is more likely than humans to form biases when hiring
MIT Technology Review Michelle Kim
AI models playing a hiring game invented their own stereotypes just from watching outcomes, not from training data. They discriminated even harder than the humans in the original psychology study.
Based on reporting by MIT Technology Review, Michelle Kim — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
Researchers at Princeton and the University of Chicago built a simple hiring simulation to see how large language models behave when they get to learn from experience instead of just parroting patterns from training data. They put ChatGPT, Claude, Gemini and others in the role of a city consultant filling 20 jobs, from doctors to janitors, choosing among candidates from four made-up ethnic groups. Every candidate was equally qualified. The models didn't know that.
Within a handful of rounds, the models started sorting people by group anyway. One bad outcome for a fictional "Aima" doctor was enough to push a model toward funneling that entire group into lower-status jobs like janitorial work for the rest of the 40-round game. Humans playing a similar game decades ago in the original psychology study did this too, scoring 0.84 on a segregation scale that tops out at 2. The AI models blew past that. OpenAI's o3 hit 1.83, nearly maxing out the scale, and models overall scored about 65% worse than the humans.
Princeton PhD student Ryan Liu, a coauthor on the paper presented at ICML in July, says this comes down to what LLMs are built to do: generalize fast from thin evidence. That instinct is great for solving a math problem with two examples. It's a liability when the
My take — AI-written commentary, not fact-checked reporting
I run an AI news site, so I'm not anti-hype by nature, but this is exactly the kind of finding people building
Read more about this at: MIT Technology Review