Do AI Systems Treat Pain Differently Based on Race, Income or Housing? Mount Sinai Researchers Investigate 3.4 Million Responses

  What if AI recommends pain drugs differently based on your label? What if this is happening quietly while states like Utah approve AI in healthcare? Mount Sinai researchers ran a massive experiment. They asked 10 large AI models to recommend pain treatments. Not once. Not a thousand times. But 3.4 million times. They created 1,000 pain cases. Half were cancer. Half were non-cancer. Then they changed only demographics. Race. Income. Housing status. Gender identity. Clinical details stayed identical. For historically marginalized patients, something unexpected appeared. The models recommended more opioids. At the same time, they flagged them as higher addiction risk. AI compressed society’s contradictions into code. “They’re at risk — so prescribe more opioids.” That line reflects measurable odds ratios in the study. In cancer cases, opioid recommendations reached 79.5%. Some subgroups exceeded 84%. Yet risk scores rose in those same groups. Why does this matter now? Because this is no longer theoretical. Utah just approved AI for chronic medication renewals. OpenAI and Anthropic are building health-facing assistants. AI is moving from research labs into real systems. If bias exists inside these tools, it does not shrink at scale. It multiplies. Many people still believe AI is neutral. Many assume algorithms are objective. But models learn from society. And society contains inequality. This research is not about panic. It is about transparency. It is about evaluation. It is about safeguards before automation expands. We simplified the full paper. We included the real numbers. We linked the original research. We explain exactly what was tested and how. If AI is entering healthcare policy now, shouldn’t the public understand what researchers already found? Read the full Research Digest. Explore the podcast. See the data yourself. 👉 https://bit.ly/3OiQ9XR