Artificial intelligence tools employed in the hiring process may be more prone to bias than human recruiters, according to recent studies. Research analyzing millions of job applications has revealed that AI screening systems can systematically disadvantage certain demographic groups, particularly Black and Asian applicants. This bias can lead to a higher rate of rejection for these candidates compared to what might be expected if human recruiters were making the decisions independently.
A large-scale study examined 4 million job applications processed by an AI-based screening tool across 1,700 job postings from 150 employers. The findings indicated that 26% of Black applicants and 15% of Asian applicants applied to positions where the AI system exhibited bias against their racial group. This adverse impact was measured using the EEOC's "four-fifths rule," which flags when one group is recommended at less than 80% of the rate of the most-recommended group. The researchers calculated that if these AI systems had recommended Black and Asian candidates at the same rate as the most-favored group, approximately 40,000 more of their applications would have advanced in the hiring process.
The study also identified a phenomenon termed "systemic rejection," where candidates were rejected from multiple jobs at a rate higher than expected if companies were making independent decisions. This suggests that reliance on identical algorithms from a single vendor, creating an "algorithmic monoculture," can lead to certain individuals being broadly excluded from employment opportunities. When applicants submitted multiple applications screened by the same AI vendor, they were more likely to be rejected from every position compared to scenarios where companies made decisions independently. This contrasts with previous analyses of hiring decisions that did not focus on AI, where rejection rates were more aligned with independent company decisions.
The precise reasons for this AI-driven bias remain unclear, but researchers emphasize the need for greater transparency in how these hiring tools function. While AI was initially hoped to reduce human bias in recruitment, evidence suggests that these systems can inadvertently codify existing societal inequities. For instance, historical hiring data, which often reflects past human biases, can train AI models to perpetuate those same disparities. An example cited is Amazon's experiment with an AI recruiting tool that downgraded résumés mentioning "women's" due to training data dominated by male candidates in technical roles.
Research has explored how the design of AI algorithms can influence candidate quality and diversity. One study found that algorithms designed to "explore" rather than solely predict based on historical data can improve both the quality and demographic diversity of candidates. Typical hiring algorithms, designed to predict success based on past hires, often favor groups that have historically succeeded, thereby limiting opportunities for minorities and women.
The widespread adoption of AI in hiring, with approximately 90% of U.S. employers utilizing AI screening tools, underscores the significance of these findings. Experts advocate for independent research into algorithmic hiring and for companies to remain responsible for verifying the fairness of their AI tools and processes. Addressing AI bias requires a multi-faceted approach, including auditing training data for skewed representation, demanding transparency from AI vendors, and establishing review boards to ensure AI systems are fair, explainable, and auditable.