Evaluating gender bias in large language models in long-term care (11 August 2025)

Summary

Large language models (LLMs) are being used to reduce the administrative burden in long-term care by automatically generating and summarising case notes. However, LLMs can reproduce bias in their training data. This study evaluates gender bias in summaries of long-term care records generated with two state-of-the-art, open-source LLMs released in 2024: Meta’s Llama 3 and Google Gemma. Its findings reveal notable variation in gender-based discrepancies was observed across summarisation LLMs. 

Content

In this study, gender-swapped versions were created of long-term care records for 617 older people from a London local authority. Summaries of male and female versions were generated with Llama 3 and Gemma, as well as benchmark models from Meta and Google released in 2019: T5 and BART. Conclusions from this studies findings included:

• Llama 3 showed no gender-based differences across any metrics, T5 and BART demonstrated some variation, and the Gemma model exhibited the most significant gender-based disparities.
• Gemma’s male summaries were generally more negative in sentiment, and certain themes, such as physical health and mental health, were more frequently highlighted for men.
• The language used by Gemma for men was often more direct, while more euphemistic language was used for women.
• In the Gemma summaries, women’s health issues appeared less severe than men’s and details of women’s needs were sometimes omitted.
• While this study provides evidence of gender bias in LLM-generated summaries for long-term care, the findings are based on one specific domain and dataset. Further research is needed to assess whether similar patterns arise in other health and care settings, such as hospitals or mental health, where documentation styles and service models may differ.