Anthropic researchers argue AI must return model distributions, not single fits, to provide scientific insight

Details

The core argument is that single optimal model fits hide alternative explanations and fail to provide insight; AI should instead return the full distribution of models compatible with the data.
This approach allows scientists to identify which parameters are truly constrained by data, which remain open, and which predictions are robust across all surviving models.
The authors use the Hodgkin-Huxley model as a historical case study, noting that the 4th-power relationship in the equations was only meaningful because lower powers failed, a distinction lost in single-fit AI outputs.
Current AI systems for model building are limited to rediscovering known models; the proposed distribution-based approach aims to handle open-ended problems where no single correct answer exists.
Simulation-based inference (SBI) is highlighted as a key technique, requiring up to millions of simulations for complex models but enabling inference in seconds once the neural network is trained.
The framework extends beyond neuroscience to climate science and particle physics, where researchers already use model ensembles to map excluded vs. open theoretical versions.
AI enables modeling systems too complex for humans to write by hand, such as the Drosophila brain with 124,635 cells, shifting the focus from finding one correct model to identifying shared properties among all data-compatible models.

Key point