Bias Out-of-the-Box: An Empirical Analysis of Intersectional Occupational Biases in Popular Generative Language Models

TL;DR

This study probes biases in out-of-the-box generative language models by examining GPT-2’s occupational associations across gender and multiple protected attributes, using a template-based pipeline to generate 396K sentences. It quantifies intersectional effects with logistic regression across 262 models and anchors predictions to US labor market data, revealing both reflect-and-correct dynamics relative to real-world distributions. The findings show gender-based clustering and significant intersectional effects, with predictions often aligning with ground-truth patterns but sometimes exaggerating stereotypes, underscoring normative questions about what these models should learn. The work highlights practical implications for deploying widely-used language models in hiring-related tasks and calls for broader, cross-country fairness analyses.

Abstract

The capabilities of natural language models trained on large-scale data have increased immensely over the past few years. Open source libraries such as HuggingFace have made these models easily available and accessible. While prior research has identified biases in large language models, this paper considers biases contained in the most popular versions of these models when applied `out-of-the-box' for downstream tasks. We focus on generative language models as they are well-suited for extracting biases inherited from training data. Specifically, we conduct an in-depth analysis of GPT-2, which is the most downloaded text generation model on HuggingFace, with over half a million downloads per month. We assess biases related to occupational associations for different protected categories by intersecting gender with religion, sexuality, ethnicity, political affiliation, and continental name origin. Using a template-based data collection pipeline, we collect 396K sentence completions made by GPT-2 and find: (i) The machine-predicted jobs are less diverse and more stereotypical for women than for men, especially for intersections; (ii) Intersectional interactions are highly relevant for occupational associations, which we quantify by fitting 262 logistic models; (iii) For most occupations, GPT-2 reflects the skewed gender and ethnicity distribution found in US Labor Bureau data, and even pulls the societally-skewed distribution towards gender parity in cases where its predictions deviate from real labor market observations. This raises the normative question of what language models should learn - whether they should reflect or correct for existing inequalities.

Figures (31)

• Figure 1: Summary table of data collection showing the number of calls per category and per variant (Var). The total number of calls is 396,000.
• Figure 1: Data Collection Process. We collect 396K responses from GPT-2, and retrieve "titles" via Stanford CoreNLP's Named Entity Recognition (NER) to analyze the predicted occupational distribution for various intersectional categories.
• Figure 2: GPT-2 occupational stereotyping. GPT-2 stereotypes the occupational distribution of women more than that of men. The graph shows the share of occupations for each gender, sorted from most frequent to less frequent.
• Figure 2: Gini coefficients of rank-frequency distributions returned by GPT-2.
• Figure 3: Fundamentally skewed GPT-2 output distributions. We show the gender proportions when querying for the base case, i.e. $X=\{\}, Y = \{\hbox{Man}, \hbox{Woman}\}$ and present all jobs with greater than $35=n*0.25\%$ mentions, making up 81% of returned sentence prompts.