Earlier this year, the UK-based law firm Addleshaw Goddard encountered a new challenge while recruiting for an IT position within its human resources department. The firm received 178 applications—far more than in previous years—reflecting the growing ease with which candidates can produce résumés using artificial intelligence (AI) tools.

To manage the influx, the firm’s HR team experimented with AI-assisted screening using its internal Microsoft Copilot system. After human reviewers narrowed the applicant pool to about a dozen, the AI agent was tasked with ranking these candidates based on criteria including experience with HR systems, programming languages, and AI familiarity. The outcome, however, revealed a divergence: the AI’s top picks did not align with those preferred by human evaluators.

The AI system tended to favor applications that included embellished claims of rapid, significant IT achievements or involvement in major projects despite limited experience—indicators suggestive of résumés crafted with AI assistance rather than genuine credentials. Katherine Rathbone, head of HR at Addleshaw Goddard, noted that the AI’s focus on candidates’ AI skills might have led the system to interpret signs of AI-generated résumés as proof of proficiency with AI itself.

This phenomenon mirrors findings from a working paper that analyzed AI hiring tools across 24 occupations, including teachers, consultants, chefs, and engineers. Researchers found that candidates who used the same AI large language model as the recruiting tool were between 23% and 60% more likely to be shortlisted compared to applicants with human-written résumés. The bias was most pronounced in business roles such as sales and accountancy, and less marked in sectors like farming and automotive work.

Jiannan Xu, one of the paper’s authors, has warned that such AI systems may exhibit “self-preferencing” behavior, potentially favoring candidates whose applications reflect AI-generated content. This raises concerns amid the expanding use of AI in recruitment, where candidates increasingly rely on AI to enhance or create their applications, intensifying competition and complicating fairness.

However, the study also points to possible mitigation strategies. Researchers suggest that instructing AI models to disregard the origin of résumés and concentrate solely on substantive content, alongside deploying multiple assessment models, can reduce bias by more than half.

At Addleshaw Goddard, the final hiring decision remained with human assessors, who selected a candidate different from the one ranked highest by the AI agent. Rathbone emphasized that while AI tools hold promise for improving recruitment efficiency, human judgment continues to be essential in making fair and accurate hiring decisions. The experience underscores the evolving complexities posed by AI integration in recruitment processes and the need for careful oversight.