"Kelly is a Warm Person, Joseph is a Role Model": Gender Biases in LLM-Generated Reference Letters

Abstract: LLMs have recently emerged as an effective tool to assist individuals in writing various types of content, including professional documents such as recommendation letters. Though bringing convenience, this application also introduces unprecedented fairness concerns. Model-generated reference letters might be directly used by users in professional scenarios. If underlying biases exist in these model-constructed letters, using them without scrutinization could lead to direct societal harms, such as sabotaging application success rates for female applicants. In light of this pressing issue, it is imminent and necessary to comprehensively study fairness issues and associated harms in this real-world use case. In this paper, we critically examine gender biases in LLM-generated reference letters. Drawing inspiration from social science findings, we design evaluation methods to manifest biases through 2 dimensions: (1) biases in language style and (2) biases in lexical content. We further investigate the extent of bias propagation by analyzing the hallucination bias of models, a term that we define to be bias exacerbation in model-hallucinated contents. Through benchmarking evaluation on 2 popular LLMs- ChatGPT and Alpaca, we reveal significant gender biases in LLM-generated recommendation letters. Our findings not only warn against using LLMs for this application without scrutinization, but also illuminate the importance of thoroughly studying hidden biases and harms in LLM-generated professional documents.

• The paper shows that LLMs produce reference letters with inherent gender biases by analyzing both minimal and detailed prompt scenarios.
• It employs social science frameworks to quantify lexical and language style biases, highlighting preferential adjectives linked to gender.
• The study introduces 'hallucination bias,' where unsubstantiated, stereotype-amplifying content is generated, impacting fairness in professional evaluations.

Gender Biases in LLM-Generated Reference Letters: A Critical Analysis

The paper "Kelly is a Warm Person, Joseph is a Role Model: Gender Biases in LLM-Generated Reference Letters" presents an in-depth investigation into the gender biases manifest in reference letters produced by LLMs, specifically exemplified by models such as ChatGPT and Alpaca. This study is critically important, as it addresses the significant and often overlooked issue of fairness and bias in automated text generation, with profound implications for real-world professional scenarios, including hiring and admissions processes.

Methodological Approach

The researchers identified two primary scenarios for evaluating bias in LLM-generated reference letters: Context-Less Generation (CLG) and Context-Based Generation (CBG). In CLG, the model generates letters based solely on minimal input, such as a name and gender, to evaluate inherent biases. CBG offers a more complex prompt, incorporating personal and professional biographical details, simulating a realistic application of LLMs where users provide comprehensive information.

The study draws upon social science frameworks to assess biases along two dimensions: language style and lexical content. It also considers "hallucination bias," a novel concept introduced here, which refers to bias exacerbation in the non-entailment, generated content that deviates from factual input, revealing how LLMs might amplify existing stereotypes during text generation.

Key Findings

The paper's findings unequivocally reveal persistent gender biases across both CLG and CBG scenarios:

1. Lexical Content Biases: The odds ratios calculated show a significant inclination in word choice reflecting gender stereotypes—male names are associated with terms like "leader" or "genius," whereas female names align more frequently with words like "interpersonal" or "warm." These biases echo prevalent societal stereotypes documented in psycholinguistic studies.
2. Language Style Biases: Analysis shows that LLMs engender reference letters where male candidates are described with more formal, positive, and agentic language compared to their female counterparts. For example, language describing men often aligns more closely with traits valued in professional settings, like assertiveness and leadership.
3. Hallucination Bias: The analysis of inconsistencies indicates that hallucinated content often exacerbates existing gender biases, where additional unsubstantiated details skew gender representation unfavorably, further entrenching stereotypes.

Implications and Future Directions

The implications of these biases in LLM-generated texts are far-reaching. The automated generation of biased reference letters can influence critical decisions in employment and academia, potentially perpetuating gender inequality without conscious intervention. Recognizing and mitigating these biases is crucial in ensuring fair and equitable AI tools.

Future work in this field must focus on developing frameworks to correct biases in LLM outputs. This includes enhancing dataset diversity, refining model training techniques to diminish bias emergence, and potentially incorporating bias-check mechanisms during generation. Additionally, extending this research paradigm to other demographic intersections, such as race or ethnicity, would provide a broader understanding of representation biases in LLMs.

The paper effectively highlights the need for critical engagement with AI technologies tasked with generating professional documentation. While LLMs present enormous potential in automating writing, the ethical and societal consequences of their inherent biases cannot be ignored. Consequently, this study calls for more rigorous academic and policy-driven discourse to navigate the implications of AI in society responsibly.