As artificial intelligence (AI) becomes a prevalent tool in the hiring process, it raises critical questions regarding fairness and bias in candidate selection. Recent research from Princeton University and the University of Chicago highlights a troubling tendency of large language models (LLMs) to form biases in hiring scenarios, potentially more so than human recruiters. These models, which include well-known systems like ChatGPT, Claude, and Gemini, were tested in a simulated hiring game designed to mimic human stereotype formation, revealing that they can develop their own biases based on their experiences.
In this study, LLMs were tasked with hiring for various jobs, including doctors and lawyers, using candidates from fictional ethnic groups. Despite equal success rates among all candidates, the models quickly began to categorize applicants based on early hiring outcomes. For instance, if an Aima candidate was unsuccessful as a doctor—a role deemed to require warmth and competence—the models would avoid hiring Aimas for similar positions in the future, instead relegating them to jobs perceived as lower in competence. This pattern persisted, with the models displaying a stronger inclination to stereotype than human participants in the original study. Such findings indicate that LLMs are adept at generalizing from limited data, often leading to premature and biased conclusions.
Moreover, the study suggests that the sophistication of the LLMs plays a significant role in their tendency to stereotype. More advanced models demonstrated even greater biases, as they were trained to prioritize optimization of outcomes based on historical data. This propensity to generalize can be problematic in social contexts; when models make quick decisions based on prior experiences, they risk perpetuating biases. Researchers noted that while instructing these models to prioritize fairness had minimal impact on their behavior, incentivizing diversity in hiring resulted in a notable reduction of biases. This implies a need for the design of AI systems that align their operational goals with socially desirable outcomes. As companies increasingly integrate LLMs into hiring processes, understanding the implications of these biases will be crucial to ensuring fair employment practices.
Source: AI is more likely than humans to form biases when hiring via MIT Technology Review