Biased Oracle: Empathy & Understandability in LLM-Generated Medical Explanations
Shunchang Liu, Guillaume Drui, Jianzhou Yao, Rikard Pettersson
Human-Centered AI for Social Good: Peace, Health, Climate, ETH Zürich
Code ·
Abstract
Large language models can tailor diagnostic explanations and tone to patient context, but do they remain understandable and empathetic across demographics? Biased Oracle evaluates GPT-4o and Claude-3.7 on 156 scenarios spanning demographics, education, conditions, and age. We combine readability metrics and LLM-as-a-Judge with human validation to expose systematic bias patterns.
Key Findings
- Diagnosis dominates empathy (Alzheimer’s highest; heart disease lowest).
- Education bias (medical degree → lower affective empathy).
- Age effects are rater-dependent (U-shape visible with some evaluators).
- LLM-as-Judge disagreement (poor inter-rater reliability between models).
Quick Start
```bash pip install -r requirements.txt