Large language models (LLMs) show opioid prescribing bias. The article was published in Nature Medicine online on February 27, 2026 with doi:10.1038/d41591-026-00013-7. A related study evaluated the bias and fairness of an NLP-based opioid abuse classifier. The classifier achieved a sensitivity of 80% (95% CI: 77–83%) and a specificity of 99% (95% CI: 99–99%) when externally validated on the first 24 hours of clinical notes. Opioid-abusing patients were disproportionately younger, male, and black compared to nonabusing patients, more likely to be on Medicaid, and equipped against a medical recommendation. The local LIME surrogate model used had an average median R² of 0.979 (IQR 0.972–0.984), which represents an excellent approximation of the predictions. Efforts to reduce bias have increased the predicted positive rate for black patients, which may lead to overtreatment and increased stigma.