Austrian study finds AI recruiting tools do not reduce hiring bias

A master’s thesis by Maria‑Elisa Michna examined whether AI‑based evaluation systems can lower unconscious bias in the early stage of personnel selection. The experimental design used 120 realistic job applications arranged in 50 paired sets that differed only in gender, ethnic origin (signaled by name), age or educational institution. A total of 160 assessments were carried out by four human recruiters and three AI systems – ChatGPT‑4o, ChatGPT‑5.1 and Google Gemini Flash 2.5 – using standardized scorecards and a seven‑step rating scale.

Bias was measured as the mean difference between the paired profiles. Human reviewers showed no statistically significant directed bias across the four dimensions studied. In contrast, the AI systems displayed heterogeneous, model‑ and feature‑dependent bias patterns, most prominently an age bias that favoured younger candidates. Direct comparisons revealed no evidence that AI reduced bias; in several cases the AI tools exhibited stronger biases than the human assessors. A repeat analysis across two rating rounds confirmed these patterns. The study concludes that deploying AI in recruiting does not automatically lead to fairer decisions; biases remain context‑, model‑ and attribute‑specific, requiring controlled use, ongoing evaluation and a human‑in‑the‑loop approach.