AI Bias in Hiring Influences Human Decision-Makers, UW Study Finds

Human Reviewers Mirror AI Bias

A new University of Washington study examined how people make hiring decisions when supported by biased AI recommendations. In a controlled experiment, 528 U.S. participants evaluated candidates for 16 jobs, ranging from nurse practitioner to computer systems analyst. Each candidate pool included equally qualified white, Black, Hispanic, and Asian men, along with one unqualified distractor candidate.

When participants received no AI input or neutral recommendations, they selected white and non-white candidates at similar rates. But once the LLM’s recommendations displayed even moderate racial bias, participants adopted the same bias. When the AI favored white candidates, so did participants; when it favored non-white applicants, participants mirrored that preference. Under conditions of severe bias, human decisions aligned with AI recommendations nearly 90% of the time.

The findings were presented at the 2025 AAAI/ACM Conference on Artificial Intelligence, Ethics, and Society.

Experimental Design

Participants were recruited via Prolific and assigned four rounds of hiring decisions:

• No AI recommendations
• Neutral AI recommendations
• Moderately biased AI recommendations
• Severely biased AI recommendations

Candidate race was implied through names, such as Gary O’Brien, and affinity group memberships. The resumes were generated using AI and validated for quality. To precisely control for bias, the research team simulated the AI recommendations rather than relying on live model outputs.

Bias Persists Even When Recognized

Although participants could sometimes detect severe bias, awareness alone was insufficient to correct their decisions. “Unless bias is obvious, people were perfectly willing to accept the AI’s biases,” said lead author Kyra Wilson.

Senior author Aylin Caliskan added that restricted access to real hiring data necessitated lab-based methods, which still yielded actionable insights into human-AI interaction.

Potential Mitigations

The study identified two promising interventions:

• Implicit association tests: Starting each session with an IAT reduced bias by 13%.
• AI literacy training: Teaching users how AI models work increased awareness of limitations and reduced overreliance on them.

However, researchers emphasized that responsibility should not be placed solely on end-users. Model developers must address bias in design, and policymakers must establish guardrails to ensure alignment with societal standards.

Broader Implications

The results underscore the risks associated with integrating AI into high-stakes decisions without adequate oversight. As long as organizations continue to use AI in hiring, often without rejecting candidates before human review, biased recommendations can systematically distort outcomes.

Co-authors include Anna-Maria Gueorguieva and Mattea Sim. The U.S. National Institute of Standards and Technology funded the study.