Professor Tristan Botelho of the Yale School of Management has identified that large language models can incorporate racial and gender biases when evaluating job candidates. The study describes a phenomenon termed symbolic norm compliance, where models produce neutral-sounding text while their internal assessments remain skewed by demographic signals.
The research highlights a significant challenge for companies using AI in hiring. While some developers implement post-training safety alignment to penalize biased responses, the study questions whether this process improves fairness or merely masks underlying issues to avoid public scrutiny. This contrasts with earlier algorithmic failures, such as the 2018 Amazon case where a recruiting tool penalized resumes containing the word 'women'.
Modern language models are now capable of bypassing simple keyword filters while maintaining biases within their internal weights. The authors argue that auditing these systems requires more than checking for prohibited language. Instead, they suggest conducting sensitivity analyses and comparing outcomes across different demographic groups to ensure genuine fairness in automated recruitment processes.
