The fact Researchers at Princeton University and the University of Chicago have found that large language models (LLMs) are more likely than humans to develop biases in simulated hiring scenarios. The study, covered by MIT Technology Review on July 20, 2026, reveals that AI does not merely reflect stereotypes embedded in training data — it can generate entirely new prejudices on its own, and at a faster rate than real people. The findings challenge one of the core promises of algorithmic hiring: that machines would bring impartiality to recruitment decisions.
Context The experiment used a simulated hiring game in which both human participants and LLMs evaluated resumes and made selection decisions. While conventional wisdom held that AI systems would be more objective since they carry no emotions or personal history, results showed the opposite: models formed stereotypes about candidates more quickly than human participants did, and then applied those biases consistently across rounds. The findings arrive at a time when tech companies are racing to build agentic models that remember granular user details — which, the authors warn, could supply even more ammunition for algorithmic prejudice to emerge in real-world hiring pipelines.
Analysis This outcome directly challenges one of the most widely held assumptions about AI in human resources — namely, that algorithms are inherently fairer than human evaluators. If machines can not only reproduce existing biases but generate novel ones through experience, the case for objectivity collapses entirely. According to MIT Technology Review's coverage, the phenomenon underscores the urgent need for new behavioral-audit protocols for AI systems — not just data-centric bias checks, but continuous monitoring of how models evolve their decision-making criteria over time. The research also raises questions about the growing trend toward agentic AI systems that learn from each interaction.
What to watch Regulators in the European Union and the United States are already debating whether to include dynamic bias testing in future AI certification frameworks for hiring. Companies relying on LLM-powered recruitment tools should take note: the absence of explicit bias in training data is no guarantee of impartiality in practice. The study suggests that one-off audits are insufficient — continuous behavioral monitoring of deployed models will likely become the new gold standard for AI governance in HR. The open question remains: if AI develops biases faster than humans do, are we prepared to audit it at the same speed?
Source: MIT Technology Review
