Investigating Gender Stereotypes in Large Language Models via Social Determinants of Health

Trung Hieu Ngo; Adrien Bazoge; Solen Quiniou; Pierre-Antoine Gourraud; Emmanuel Morin

cs.CLMar 10, 2026

Investigating Gender Stereotypes in Large Language Models via Social Determinants of Health

Authors: Trung Hieu Ngo, Adrien Bazoge, Solen Quiniou, Pierre-Antoine Gourraud, Emmanuel Morin

Organizations: Nantes Université, École Centrale Nantes, CNRS, LS2N, UMR 6004, F-44000 Nantes, France · Nantes Université, CHU Nantes, Clinique des données, INSERM, CIC 1413, Nantes, France

Abstract

Large Language Models (LLMs) excel in Natural Language Processing (NLP) tasks, but they often propagate biases embedded in their training data, which is potentially impactful in sensitive domains like healthcare. While existing benchmarks evaluate biases related to individual social determinants of health (SDoH) such as gender or ethnicity, they often overlook interactions between these factors and lack context-specific assessments. This study investigates bias in LLMs by probing the relationships between gender and other SDoH in French patient records. Through a series of experiments, we found that embedded stereotypes can be probed using SDoH input and that LLMs rely on embedded stereotypes to make gendered decisions, suggesting that evaluating interactions among SDoH factors could usefully complement existing approaches to assessing LLM performance and bias.

Figures & tables

Figure 1: A sample of input data override in a diagnosis task. Although the gender was specified as Male in the input data, the model continues to suggest menstrual-related problems as one of the diagnoses.

Figure 2: Averaged number of occurrences for each predicted class across 3 runs for each model. Error bars indicate Standard Deviation values.

Figure 3: Modified RMSE scores for 9 models, denoting the deviation from the neutral value of 4. A positive score means an overall preference for Masculine predictions and vice versa. Larger absolute values show higher bias degrees.

Figure 4: Heatmap of associations between SDoH options and the Male (left) and Female (right) predictions for nine models. Odds ratio values are reported in the cells. Color intensity indicates probability. Statistically significant odds ratio values are marked with an asterisk (*).

Figure 5: Heatmap of associations between Profession groups and the Male (left) and Female (right) predictions for nine models. Odds ratio values are reported in the cells. Color intensity indicates probability. Statistically significant odds ratio values are marked with an asterisk (*).

Figure 6: Prediction tendencies of 9 annotators.

Figure 7: Association heatmap between Profession groups and Male (top) and Female (bottom) predictions of LLMs and annotators on 100 examples. Odds ratio values are reported in the cells. Color intensity indicates probability. Statistically significant odds ratio values are marked with an asterisk (*).

Models	Full	Filtered	Extracted	Neutr.
	text	text	SDoH	SDoH
Llama3.1-8b	2.51	2.54	2.60	2.25
Llama3.3-70b	2.48	2.48	2.35	1.56
Qwen2.5-7b	2.34	2.35	2.35	0.97
Qwen2.5-72b	1.88	1.89	1.85	0.96

Table 1: Modified RMSE scores for different input data formats, tested on 100 examples.

Appendix figures & tables7 assets

Supplementary material from the paper’s appendix.

Appendix

SDoH	Options	Description
Living condition	Alone	The patient lives alone.
	With others	The patient lives with other people.
Marriage status	Single	The patient is described as not in a relationship.
	Married/In relationship	The patient is described as in a relationship with a partner or in a marriage.
	Divorced	The patient is described as separated from a partner or divorced.
	Widowed	The patient is described as widowed.

Appendix

Table 2: Social Determinants of Health (SDoH) with Options and Descriptions. SDoH without Options are annotated using span-only labels.

(French) Free-form text	Transformed structured input
Mode de vie: Marié, vit avec sa femme dans une maison avec un étage.	Conditions-de-vie_Cohabitation: Oui; Statut-matrimonial_Marié: Oui; Domicile: Oui;
Pas d’aide à domicile, autonome.
Une fille de 57 ans, et un fils de 53 ans dans le sud Est.	Descendance: Oui;
Retraité, ancien directeur d’usine.	Statut-emploi_Retraité: Oui; Dernière-profession: Directeur d’usine/Directrice d’usine;
Pas d’exogènose rapportée par son épouse.	Consommation-alcool_Non: Oui;
Marche avec un déambulateur a priori.

Appendix

Table 3: An example of transformed input in French and translated to English for readability. The gendered information presented in the “Living condition” and “Last occupation” SDoH was neutralized in the French version to ensure the least amount of bias in the input data.

Panel of Figure 8 — Figure 8: Input prompt in French for gender prediction. The models are also provided with the full list of possible SDoH to ensure comprehension of input data.

Panel of Figure 9 — Figure 9: Input prompt for gender prediction (Translated English version). The models are also provided with the full list of possible SDoH to ensure the comprehension of input data.

Free-form text	Filtered text
Mode de vie: Vit à domicile avec sa femme, autonome pour les activités de la vie quotidienne.	Mode de vie : Vit à domicile avec sa femme, autonome pour les activités de la vie quotidienne.
Trois enfants.	Trois enfants.
Ancien chercheur en biochimie en retraite.	Ancien chercheur en biochimie en retraite.
Nombreux voyages notamment Sénégal vers 1967, Thaïlande, Tunisie vers 1957, Madère en 2013.
Tabagisme à 60 paquets année (1 paquet par jour pendant 60 ans) sevré.	Tabagisme à 60 paquets année (1 paquet par jour pendant 60 ans) sevré.
Consommation d’alcool 10 à 20 g/jour max.	Consommation d’alcool 10 à 20 g/jour max.

Appendix

Table 4: An example of the four different formats of input in French.

Figure 10: Modified RMSE scores of 9 annotators. Those relying on stereotypes for decision-making are annotators 2, 6, 7, 8, 9, and those favoring neutral judgments are annotators 1, 3, 4, 5. Modified RMSE scores for the more neutral group are generally smaller than the other group, except for Annotator 3.

Figure 11: Associations between Male/Female predictions and SDoH options. Statistically significant values are marked with an asterisk (*)

Intersectional Bias in Language Models

See all 9 papers

Gender bias across LLMs is common and highly heterogeneous
Sep 29, 2026Edoardo Bolzoni, Valerio CapraroGender Bias in Language Models LLM Auditing
University of Milan-Bicocca, Milan, Italy
Artificial Intolerance: Stigmatizing Language in Clinical Documentation Skews Large Language Model Decision-Making
May 17, 2026Jen-tse Huang, Didi Zhou, Faith Kamau +5Social Bias in Language Models Clinical Language Model Evaluation
1Johns Hopkins University.
Anchoring LLM Gender Bias to Human Baselines: A Cross-Lingual Audit
May 29, 2026Jiwoo Choi, Seonwoo Ahn, Tongxin Zhang +1Gender Bias in Language Models Multilingual Language Model Evaluation
School of Digital Humanities and Computational Social Sciences, KAIST, South Korea

Sep 29, 2026J/K move · Enter open · S save

Gender bias across LLMs is common and highly heterogeneous

Edoardo Bolzoni, Valerio Capraro

University of Milan-Bicocca, Milan, Italy

Gender Bias in Language Models LLM Auditing Language Model Bias Evaluation

Open paper AskPDF arXiv

Understanding gender biases in large language models (LLMs) is increasingly important as these systems become embedded in decision-support tools with real consequences. Prior research has focused only on a small set of models, leaving open the extent to which gender biases are common and heterogeneous across LLMs. We address this gap across ten models released between April 2025 and June 2026, spanning nine vendors, using two paradigms: gender attribution to stereotyped phrases (Study 1) and moral judgment of abuse or torture against a woman or a man to prevent a catastrophic outcome (Study 2). In Study 1, two of ten models attributed masculine-stereotyped phrases to female writers more often than the reverse, while three models showed the opposite pattern. In Study 2, several models converged on a male-disadvantaging asymmetry that was directionally consistent with a documented human tendency to protect female targets from harm, though the specific conditions under which this asymmetry emerged varied by model; three other models, by contrast, showed no variation across conditions. These results indicate that gender-related biases are common in LLMs. Their direction and magnitude, however, are highly heterogeneous, to the point that some models behave in diametrically opposite ways to others. Bias auditing should therefore be treated as an ongoing, multi-vendor process, rather than a one-time assessment.

Abstract

Intersectional Bias in Language Models

Explore similar work

Gender bias across LLMs is common and highly heterogeneous

Artificial Intolerance: Stigmatizing Language in Clinical Documentation Skews Large Language Model Decision-Making

Anchoring LLM Gender Bias to Human Baselines: A Cross-Lingual Audit