Patient-facing medical question answering with large language models is moving from research demonstrations to production deployments, according to a new paper on arXiv. The authors argue that current evaluation practices remain anchored on factual correctness, but that factuality alone is not enough for safe deployment.
The paper's central claim is that medical fine-tuning does not simply reduce risk; it reshapes it. That means risk assessment must look beyond whether an answer is factually right and consider other dimensions of harm that fine-tuning may introduce or alter.
To support such assessment, the authors propose a rubric-based evaluation framework that uses an LLM judge to score outputs along broader criteria. The framework is intended to give developers and regulators a more complete picture of patient-facing AI risk than factuality benchmarks alone.