Reducing Bias in HR With AI: What the Research Says
The claim that AI reduces bias in HR is one of the most frequently made and most frequently misunderstood promises in HR technology marketing. The reality is more nuanced: AI can reduce certain types of bias, but it introduces new ones, and whether the overall effect is a reduction in bias depends entirely on how the AI is designed, trained, and monitored. Understanding the research is essential for making informed decisions about where AI helps and where it makes the bias problem worse.
The types of bias that AI reliably reduces are those driven by human fatigue, inconsistency, and in-group favouritism in high-volume tasks. When a human screener reviews two hundred CVs over an afternoon, their criteria drift: the first twenty applications receive careful assessment; by application one hundred and fifty, fatigue has introduced systematic patterns that have nothing to do with candidate quality. AI screens application two hundred with the same criteria as application one. This consistency advantage is real and valuable.
The types of bias that AI reliably introduces are those embedded in training data and problem formulation. The most documented example in HR is recruitment screening: when AI models are trained on historical hiring data, they learn patterns that reflect not just job performance predictors but also demographic patterns in the historical hiring population. Amazon famously discontinued an AI recruitment tool in 2018 after discovering it had learned to penalise CVs that included the word "women's" — as in "women's chess club" — because women had been underrepresented in its historical hiring data for technical roles.
The research on structured versus unstructured interviews provides a useful analogy for thinking about AI in HR decisions. Structured interviews — where all candidates are asked the same questions and responses are scored against consistent criteria — produce significantly less biased hiring decisions than unstructured ones. AI that imposes structure and consistency on previously unstructured decisions (like CV screening) can reduce bias. AI that automates unstructured decisions (like "does this person seem like a good fit?") amplifies whatever biases were in the decisions it was trained on.
Bias monitoring is not a one-time setup exercise. A model that shows no significant demographic disparity in its outputs at deployment may develop disparity over time as the labour market, the applicant pool, or the organisation's workforce changes. Regular auditing of AI outputs against demographic data — quarterly at minimum for high-volume AI decision tools — is the ongoing quality management that responsible implementation requires.
The most important finding from the research is that AI can be a powerful tool for bias reduction when deployed in specific, well-monitored applications: structured interview scoring, CV screening against objective criteria, pay equity analysis, and performance rating calibration across teams. It is a liability for bias when deployed in less structured applications without monitoring: personality assessment, cultural fit evaluation, and sentiment-based attrition prediction are areas where the research shows mixed to negative effects on bias.
Mellow's AI tools are designed around the applications where AI bias reduction is well-evidenced: structured screening criteria, pay equity analysis, and performance data aggregation. For applications where bias risk is higher, Mellow routes to human judgement rather than AI decision-making. The platform maintains demographic disparity reporting on all AI-assisted outputs, so that HR teams can see whether the tools they are using are producing equitable outcomes and take corrective action if they are not.