AI has a big workplace bias problem

TL;DR: Unfortunately, the AI that (probably) looked over your job application is biased—maybe even more than humans are. New research shows that LLMs don’t just parrot human prejudices, but can invent new ones all by themselves—a worrying finding as some 90% of companies already use AI in their hiring process.

What happened: As part of a hiring simulation, researchers at Princeton and the University of Chicago set AI models loose on filling 40 positions, each time choosing one of four candidates from four fake ethnic groups. The result: The AI pigeonholed the groups more aggressively than humans did. (This study is modeled after a similar one that explored bias in humans.)

LLMs scored 65% higher than humans did on job segregation—basically, how strongly each ethnic group got funneled into a single job type, whether that’s a doctor, politician, waiter, or garbage collector. For example, after an AI model was told that someone it hired didn’t make a good doctor (a random outcome), it avoided hiring people from that group overall as doctors, instead giving them jobs as janitors.

The problem, according to the researchers, is that LLMs are built to generalize even when they have limited data. “That’s literally a lot of what they’re optimized for,” one of the paper’s coauthors told MIT Technology Review.

Bigger brains, more bias: A range of frontier models were put to the test (the newest ones at the time being GPT-5.4, Claude Sonnet 4.6, and Gemini 3 Flash). Newer, more capable models stratified ethnic groups by job type even more than their predecessors, because they drew “more precise inferences” from just a few past outcomes, per researchers.

This tendency to overgeneralize might only get worse as AI models get better memory, potentially leading them to “over-index” on past behavior, one expert told MIT Tech Review.

AI bias IRL: AI is already being used in hiring, but it may increasingly be a part of weighing employee performance, too. Meta’s come under fire for allegedly tracking employees’ AI token use, keystrokes, and computer activity to score their productivity—then leaning on those metrics to determine who to lay off, according to a new lawsuit filed by Meta employees last week. The suit claims the layoffs disproportionately targeted people on medical or parental leave who couldn’t rack up productivity scores.

Bottom line: It’s unlikely that companies are going to stop using AI to judge current and prospective employees—and so far, workplaces haven’t been very transparent about when and how they use AI for such performance and hiring insights. —WK