Assessing Racial and Ethnic Bias in Text Generation by Large Language Models for Health Care-Related Tasks:...

Area of Science:

• Artificial Intelligence in Healthcare
• Natural Language Processing
• Health Equity

Background:

• Growing concern over racial and ethnic bias in large language models (LLMs) used in healthcare.
• LLM safeguards exist against overtly biased prompts, but subtle biases remain a concern.

Purpose of the Study:

• Investigate potential racial and ethnic bias in four LLMs (GPT-3.5-turbo, GPT-4, Gemini-1.0-pro, Llama3-70b).
• Assess bias in generating healthcare consumer-directed text without overtly biased queries.

Main Methods:

• Cross-sectional study prompting four LLMs to generate HIV discharge instructions.
• Varied patient race/ethnicity (African American, Asian, Hispanic White, non-Hispanic White) in deidentified metadata.
• Analyzed LLM output for sentiment, subjectivity, reading ease, and word frequency.

Main Results:

• No statistically significant differences in linguistic or readability measures across racial/ethnic groups.
• A minor significant difference in entity count for GPT-4 was not supported by post hoc analysis.

Conclusions:

• Four evaluated LLMs demonstrated relative invariance to race/ethnicity in healthcare text generation.
• Urgent need for standardized methods to measure bias in LLM-generated health content.
• Further validation and implication studies are required.