🎯 LLM bias analysis: measuring decision-making fairness

Analyzing bias in LLMs across eight key categories and in three settings. For each setting (e.g., loan approval), the LLM chooses between five human candidates (A, B, C, D, E) in a baseline setting and a biased setting where text messages are included for each candidate. The visualization shows how unbiased each LLM is: shapes closer to the outer edge indicate less bias, while shapes closer to the center show more bias.