Job seekers are facing a new challenge in the hiring process due to the rise of AI-powered applicant tracking systems (ATS). A recent study reveals that these systems tend to favor resumes generated by AI over those written by humans. Notably, candidates using the same large language model (LLM) as the company’s ATS are more likely to be shortlisted, creating a problematic bias in hiring practices.
The research, conducted by scholars from the University of Maryland, the National University of Singapore, and Ohio State University, indicates that this bias could lead to the exclusion of qualified candidates who do not utilize the same AI tools. The authors emphasize that if left unaddressed, this phenomenon could distort hiring outcomes, unfairly favoring specific applicants and potentially overlooking strong candidates who produce human-written resumes.
In examining over 2,200 human-written resumes, the researchers also constructed various counterfactual versions using advanced LLMs. They simulated hiring processes for 24 different occupations and discovered that automated evaluators were 23% to 60% more likely to select candidates who had utilized the same language model as their evaluation system. The issue was particularly pronounced in fields such as accounting, sales, and finance.
Experts in the field, including Boston University’s Professor Emma Wiles, argue that such self-preferencing undermines the objective of AI in recruitment, which is to assess a candidate’s capabilities fairly. Wiles urges job seekers to leverage AI tools that enhance their writing rather than replace their original voice, underscoring the importance of authenticity in applications.
Why this story matters
- The study highlights a new form of bias in the hiring process, impacting equitable job opportunities.
Key takeaway
- AI evaluations tend to favor resumes created by the same model used by employers, potentially sidelining qualified candidates.
Opposing viewpoint
- Proponents of AI argue that it streamlines the hiring process and enhances efficiency, despite the emerging biases.
