LLMs Exhibit Gender Bias in Hiring Decisions, Recommending Lower Pay for Female Candidates

The integration of large language models into high-stakes decision-making processes, particularly in human resources, is revealing deeply embedded societal biases that manifest in unexpected ways. A recent investigation into LLM behavior as a "hiring manager" has uncovered a paradoxical gender bias: while models are more inclined to select female candidates and perceive them as more qualified for a given resume, they simultaneously recommend lower compensation for these same candidates compared to their male counterparts. This finding exposes a nuanced and concerning form of algorithmic discrimination, where perceived competence does not translate into equitable economic valuation.

This specific bias, where female candidates are favored for selection but penalized in pay, highlights the complex and often contradictory nature of biases absorbed during LLM training. The study's use of prompt engineering as a mitigation technique underscores the ongoing challenge of de-biasing these powerful models. The fact that LLMs can simultaneously exhibit a preference for female candidates in hiring while perpetuating a wage gap suggests that simple interventions may be insufficient to address the multifaceted biases present in their training data, which reflects historical and systemic inequalities.

The implications for the future of AI in professional contexts are profound. Without rigorous and continuous auditing, AI-driven hiring tools risk codifying and amplifying existing societal inequities, leading to widespread wage discrimination and undermining efforts towards gender pay equity. This research serves as a critical warning, emphasizing that the mere presence of AI in decision-making does not guarantee objectivity or fairness. It necessitates a concerted effort from developers, policymakers, and organizations to implement robust ethical AI frameworks, transparent bias detection mechanisms, and proactive mitigation strategies to ensure that AI systems contribute to a more equitable, rather than a more biased, future workforce.