Addleshaw Goddard HR department tested an AI agent against humans to assess job applications. The AI agent, using Microsoft Copilot, evaluated 178 applications for an IT role. While humans shortlisted about a dozen candidates, the AI presented its own top contenders.
The lists did not match. The AI was impressed by candidates claiming exceptional IT feats in record time, which experts suspect were AI-tailored CVs. Katherine Rathbone, head of HR at Addleshaw Goddard, noted that since AI capability was a criterion, the agent may have interpreted signs of AI use as evidence of capability.
A working paper by Jiannan Xu and colleagues simulated hiring processes for 24 jobs. They found candidates using the same AI large language model as the recruiter were 23 to 60 percent more likely to be shortlisted. This bias was highest for business roles like sales and accountancy, and lowest for farming and automobiles.
The research suggests AI bias can be more than halved by instructing models to ignore résumé origins and focus on content, and by using multiple assessment models. This highlights a growing dilemma in recruitment as job hopefuls flood markets with AI applications, potentially stacking the deck against those using traditional methods or different AI tools.
Source: Financial Times Companies · Summarized by HeadlinesBriefing