No Racial Bias Found in AI Pain Management

🔥 What happened An audit of 17 models – GPT-4.1 mini, Claude Haiku 4.5, Gemini 2.5 Flash, Llama, DeepSeek, Qwen – found no significant race or sex bias in pain treatment decisions. All 1,386 comparisons failed to survive false-discovery-rate correction. 💡 Why it matters Models differed far more from each other and across their own settings than between patient groups: the reference patient got the high opioid dose in anywhere from 6% to 93% of responses depending on configuration. Fairness audits must measure a model's own response instability, not just between-group bias. ⚡ Our take Reassuring for anyone deploying LLMs in clinics – but the massive variance across configurations is a red flag: testing a single setting measures noise, not fairness.