Medical AI Position Paper: Treatment Outcomes Overlooked

A new position paper argues that medical AI systems are neglecting real treatment outcomes in favor of human opinions and syntheses, such as biomedical publications and clinical practice guidelines. The paper, titled 'Position: Medical AI Neglects Real Treatment Outcomes,' was submitted to arXiv (ID 2608.14598) under the Computer Science > Artificial Intelligence category. It contends that while medical AI has improved in diagnostic and prognostic tasks leading to treatment decisions, its understanding of treatment itself is inadequately trained and evaluated. The authors advocate for incorporating real treatment outcomes from observational databases and randomized experiments into both training and evaluation, reemphasizing improved outcomes as the downstream goal of all medical AI. The paper highlights deficiencies in frontier models and major benchmarks due to this neglect. The submission history is not detailed, but the paper is available on arXiv with experimental HTML, references, citations, and tools for bibliographic management. The paper is part of arXivLabs, a framework for community collaborators, and is committed to values of openness, community, excellence, and user data privacy.