Summary: On October 14, 2026, Meta shared results from an exploration into AI-assisted health research: a coding scaffold built on Muse Spark 1.3 and Muse Wonder, used by medical researchers against a database of five million de-identified electronic health records. In four weeks, the cohort produced ten research manuscripts, each reviewed by peers, a biostatistician, and independent external experts. For builders, it’s a concrete example of what a domain-expert-in-the-loop AI workflow can produce when the tooling handles the code and the humans keep the scientific judgment.
The setup: experts decide, AI writes and runs the code
The system Meta describes is explicitly human-led. AI accelerates each stage of research — hypothesis generation, code writing and execution, manuscript drafting — while domain experts remain responsible for the scientific decisions and conclusions. The researchers, with backgrounds in cardiology, epidemiology, infectious disease, and internal medicine, used the scaffold to brainstorm hypotheses, write and execute all the code, and draft manuscripts alongside the model.
They also developed what Meta calls a shared framework and skills for steering the model at each stage. That detail matters. The hard part of these systems isn’t the model; it’s the workflow conventions a team builds around it so output quality compounds instead of drifting.
What came out of it
The ten manuscripts span infectious disease, sports medicine, emergency medicine, oncology, endocrinology, oral health, and more — topics as specific as pickleball injuries, e-scooter and e-bike injuries, and breast cancer drug sequencing. Each paper went through three layers of iteration: peer feedback from the medical cohort, a biostatistician’s review of the analysis, and summary reviews from third-party contracted independent experts published alongside each paper.
One example: Ssentongo et al. found that after infective endocarditis, patients had a 73% higher rate of new arrhythmias and conduction disorders than patients with bacteremia alone, with one-year risk of seven percent versus four percent. An independent cardiologist’s review called it methodologically rigorous for observational research but “hypothesis generating rather than practice changing” — the reviewers flagged surveillance bias as a live alternative explanation and the authors declined to claim causality.
That’s a good sign, honestly. If these manuscripts were coming back glowing, I’d trust the pipeline less.
What builders should take from this
Three things stand out if you’re building AI tools for expert workflows:
- Put the expert in charge of judgment, the model in charge of mechanics. Cohort selection, variable choices, and interpretation stayed with humans. The scaffold handled code execution and drafting. That division is what made the medical reviewers willing to engage seriously with the output.
- The data mattered as much as the model. Access to a licensed, de-identified five-million-record EHR database is what made original research possible. A generic assistant with synthetic data would have produced nothing comparable.
- Build review loops into the workflow, not after it. Peer review, biostatistics checks, and external expert summaries were part of the process, which is exactly the kind of governance layer worth borrowing. It echoes the guardrail patterns I noted in Anthropic’s midterm election safeguards — checks designed before deployment, not patched in afterward.
The caveats
Meta itself frames this as an initial exploration, and notes that understanding how to use AI to accelerate health science is still early. The external reviews cutting against the flashiest findings — and the four-week timeline — suggest speed came from the scaffold, while confidence has to come from the review process. If you’re modeling a similar expert-in-the-loop system, budget for the review layers as a first-class cost, because that’s where the credibility in this example actually came from.
Meta’s full write-up, including the papers and expert reviews, is on the Meta AI Research blog.