Meta shares the human-led AI workflow behind ten health research manuscripts

The ten manuscripts are best read as a worked example of how to split observational health research between a model and the people accountable for it, not as a measurement of how well AI does research. Meta's research blog describes an internal coding scaffold built on Muse Spark 1.3 and Muse Wonder, Meta's model for agentic work, connected to public datasets and a licensed retrospective database of about five million de-identified electronic health records. Medical researchers from cardiology, epidemiology, infectious disease and internal medicine used it over four weeks to brainstorm hypotheses, write and run all of the analysis code, and draft the papers with the model, while keeping the scientific decisions and conclusions. Each manuscript then went through feedback from peers in the group, a biostatistician's review of the analysis, and summary reviews from contracted outside experts, which Meta prints under each paper.

The researchers' companion report explains why the human role is more than sign-off. It describes the model as fast and strong at drafting, extraction and code, but unreliable at judging whether a source existed, whether a recorded field measured what the study intended, or whether its own output was right, and says it was at times confidently wrong with no signal a reviewer could use. In one case it calculated a 1% mortality rate for Staphylococcus aureus bloodstream infection; an infectious disease physician knew published figures are closer to 30%, so the team stopped to audit how deaths were recorded, which revealed other implausible patterns. It also counted a small heparin flush used to keep an IV line open as the same exposure as a treatment dose, and more than once said data was missing from the dataset when it was there. The authors say the clearest gain was not faster drafting but stopping weak or already-answered questions early, and that a shared definition of the data, with checks enforcing it, kept each study from rebuilding the data structure from scratch.

The speed claims come from the participants. The blog says manuscripts went from first prompt to draft in under two weeks, some in seven days, and that researchers reported up to a tenfold improvement in the best cases. The breast cancer drug-sequencing paper is more cautious about its own workflow: research time was not prospectively measured, no human-only version was run for comparison, no quantitative efficiency gain is claimed, and AI accuracy was not benchmarked. Several of the published outside reviews also keep the clinical results provisional. One calls the arrhythmia study hypothesis-generating rather than practice-changing because surveillance bias remains a live explanation, and another says the mortality claims in the S. lugdunensis endocarditis paper are overstated. Meta says this work is limited to retrospective data and that it hopes to produce hypotheses that can later be tested in prospective trials.