Smarter AI Models Form Stronger Hiring Stereotypes, Study Finds

Employers are increasingly using large language models to screen resumes and assess candidates, betting that automation can cut costs and make hiring more consistent. A study by researchers at Princeton University and the University of Chicago, however, suggests the systems may do more than reproduce biases embedded in training data: they can develop novel stereotypes without being explicitly taught, raising concerns about fairness, accountability and compliance in automated recruitment.

The researchers compared hiring judgments made by several LLMs with those of human participants and found that the models exhibited about 65% more stereotyping overall. Bias was stronger in larger systems and those with more advanced reasoning capabilities, challenging the assumption that smarter models will necessarily make fairer decisions. The team warned that, without audits, human review and effective intervention, AI hiring tools could independently reinforce and deepen social discrimination.