Publications.

AI systems can now write research papers from start to finish. But can they actually do science? This paper investigates whether today's autonomous AI scientist systems can work with complex, real-world biomedical data and produce findings that a human expert would consider valid and meaningful.

We developed a framework that maps the full arc of scientific research — from forming a hypothesis to publishing findings — and used it to survey and classify 73 AI scientist platforms, identifying 7 that operate end-to-end with minimal human intervention. We then selected representative systems and ran them on the AI-READI dataset: real-world multimodal data spanning continuous glucose monitoring, retinal imaging, sleep, physical activity, and environmental exposures.

Each system's output was reviewed at two levels. First, does the paper hold together — is it internally consistent and logically sound? Second, does it contain real scientific value — would a domain expert consider the findings meaningful? We also compared AI-generated reviews against human expert judgment to understand where automated evaluation succeeds and where it falls short.