Measuring stereotype and deviation biases in large language models.

Area of Science:

• Artificial Intelligence
• Natural Language Processing
• Computational Social Science

Background:

• Large language models (LLMs) are increasingly integrated into various applications.
• Concerns exist regarding the inherent limitations and potential risks associated with LLM outputs.
• Understanding and mitigating bias in LLMs is crucial for responsible AI development.

Purpose of the Study:

• To investigate stereotype bias and deviation bias in advanced large language models.
• To examine how LLMs associate demographic groups with attributes like political affiliation, religion, and sexual orientation.
• To uncover biases in LLM attribute inference and their potential societal harms.

Main Methods:

• Four advanced large language models were prompted to generate individual profiles.
• The study analyzed associations between demographic groups and specific attributes within the generated content.
• Demographic distributions in LLM outputs were compared against real-world distributions.

Main Results:

• All examined LLMs demonstrated significant stereotype bias.
• All examined LLMs exhibited significant deviation bias.
• Biases were observed across multiple demographic groups and associated attributes.

Conclusions:

• Large language models exhibit demonstrable stereotype and deviation biases.
• These biases can lead to inaccurate inferences about user attributes.
• The findings highlight potential harms stemming from biased LLM-generated content.