# Scientists in an Anthropic essay say AI agents should hand back many plausible models, not one best fit

> Researchers from Tübingen argue an agent that returns a single model hides equally good alternatives. Their method weighs a whole population of candidate simulators.

- URL: https://devisdead.ai/news/scientists-in-an-anthropic-essay-say-ai-agents-should-hand-back-many-plausible-m
- Published: 2026-11-08T23:33:37Z
- Author: DevIsDead Newsroom
- Section: AI
- Tags: ai for science, anthropic, modeling, uncertainty
- Originally reported by Unite.AI (https://www.unite.ai/anthropic-guest-essay-urges-ai-to-return-ranges-of-scientific-models/).

When an AI agent builds a scientific model, should it give you the answer, or all the answers that fit? Three researchers argue for the second in a guest essay published by Anthropic, according to [Unite.AI](https://www.unite.ai/anthropic-guest-essay-urges-ai-to-return-ranges-of-scientific-models/), whose article was written by an AI research agent and reviewed by its editors. The authors, Jakob Macke, Daniel Gedon and Stefan Wahl who work at Tübingen's university and its ELLIS Institute, are grantees of Anthropic's AI for Science program.

## The argument

Reproducing data isn't the same as understanding a system. Lots of parameter combinations, and sometimes entirely different equations, fit the same data equally well. Hand back just one model that works, and the others vanish from view, which may imply different mechanisms or predictions.

Their example is Hodgkin and Huxley's model of the nerve cell. They settled on a term raised to the fourth power because lower powers matched the data less well, and that pointed to a gate in the membrane with several parts moving together. Without knowing the rejected options, the fit alone wouldn't justify the interpretation. (Huxley also spent over three weeks solving the equations on a hand-cranked calculator after Cambridge's computer broke down in 1951.)

## The method

Bayesian methods already represent uncertainty over parameters, and simulation-based inference makes that workable even when likelihoods are intractable, by running a simulator many times and training a neural estimator on the results. The authors extend it to model structure: an LLM agent writes candidate simulator code, fits it, compares it with data and revises it. The output is a weighted population of models, with overly flexible fits penalized. What the surviving models agree on is what the evidence supports; where they disagree is where the next experiment should go. They formalized this earlier this year as ModelSMC.

## Why it matters

It's a sharp counterpoint to the "AI scientist gives you the answer" pitch, and it matches what other studies this week found: agents tend to [overstate their results](/news/nine-ai-agents-tried-real-follow-up-research-for-national-lab-scientists-none-cl). An agent that returns a range, and tells you what it ruled out, is one a scientist can actually check.

**[Scoreboard](/scoreboard) update:** Draw. AI agents can build scientific models fast, and scientists want them to admit what they don't know.

Stay undead.