An Open-Source AI Research Assistant Passes Peer Review, With Seven Clinical Case Studies
A Nature Medicine paper describes OpenScientist, an AI research workflow that can search literature, write analysis code, iterate on results, and produce a report with provenance. The paper presents five primary and two supplementary case studies using clinical research data. The orchestration layer and skills library are available under Apache 2.0.
What the research demonstrates
The work focuses on auditable research outputs. Case logs let researchers compare the agent's actions, calculations, and interpretation with previously studied results. The article reports model costs below $10 for the cited runs. This figure does not include data preparation, expert review, or the costs of obtaining access to appropriate data.
Why the authors' failure cases matter
The paper reports four categories of failure, including examples in which some hypotheses still appeared statistically supported after removing a biological signal. Agreement with earlier work by itself is insufficient. Independent reviewers need to inspect the assumptions, data, and analysis trail.
What the evidence does not establish
These research case studies do not validate OpenScientist for patient care or independent clinical decisions. Its reproducible logs and open components are useful for research teams that can audit an analysis and handle clinical data under their own privacy and governance requirements.