An Anthropic guest essay published on November 6, 2026 argues that reproducing scientific measurements is not enough to explain them. For researchers, students and funders assessing AI-generated models, the practical question is which competing explanations still fit—and what evidence could separate them.
The authors—Jakob Macke, Daniel Gedon and Stefan Wahl, identified by Anthropic as AI for Science grantees—discuss research methods, not a new consumer product. Their assessment is that model-building agents have mainly been evaluated on rediscovering known results; evidence on problems with unknown answers remains sparse.
Here, a scientific model means a proposed representation of a real system, such as equations describing a nerve cell. It is distinct from the language model that may help write it.
A perfect fit can leave the answer open
Consider a mathematical illustration, not experimental evidence. Suppose two noiseless observations are (0, 0) and (1, 1). Both the straight line y = x and the curve y = x² pass through those points exactly. Their fit to the available observations is identical.
At x = 2, however, the line predicts 2 and the curve predicts 4. A measurement there could distinguish these two candidates. It would not prove that every other possible curve is wrong, nor explain a physical mechanism by itself. Real measurements also introduce noise, which can make competing predictions harder to separate.
Different settings, or a different mechanism?
A model’s parameters are its adjustable settings. In a 2020 eLife study, Pedro Gonçalves and colleagues used neural networks trained on simulations to estimate distributions of parameter values compatible with observations, rather than returning just one fitted set. Their examples included different circuit configurations producing similar rhythmic activity.
Such a distribution, called a posterior, represents uncertainty after accounting for data and starting assumptions. It does not certify that every possibility has been found. The study’s discussion notes that choices about which features of the data to analyse affect interpretation, and that unexplored parameter regions can remain.
Changing a model’s structure asks a further question: are the proposed parts and interactions themselves right? Adjusting the settings inside one mechanism does not settle whether a different mechanism could explain the observations.
ModelSMC, described in a 2026 machine-learning conference paper by Stefan Wahl and colleagues, addresses this structural uncertainty. A language model proposes and revises candidate programs; statistical calculations weight them according to their support from the data. The aim is a collection of plausible models, not simply a winning answer.
That broader search has limits. In their author preprint, the researchers identify repeated simulation costs, potentially biased statistical approximations and dependence on the language model proposing relevant structures. They describe the contribution as methodological, not proof of unrestricted scientific discovery.
A separate Model Discovery Agent preprint by Kevin Murphy makes another boundary explicit: probabilities calculated over a finite collection of candidates do not account for models never proposed. The practical implication is that confidence among shortlisted explanations is not confidence that the shortlist covers reality.
A realistic output is only one check
The sbi software documentation describes a useful diagnostic: draw parameter settings from the inferred distribution, use them to generate simulated observations, and compare those outputs with the real observations. This is a posterior predictive check. It can reveal mismatches, but the documentation cautions against treating it as a validation metric.
Other checks answer different questions. The project’s diagnostic guide distinguishes tests of whether uncertainty estimates are too narrow or too broad from checks of whether the simulator—the program representing the system—can plausibly generate the observation at all. Good uncertainty estimates within an inadequate simulator would still leave the scientific explanation in doubt.
This concern predates today’s model-building agents. In a 2022 preprint, Patrick Cannon, Daniel Ward and Sebastian Schmon reported that a mismatch between the assumed model and the process generating data could seriously degrade neural simulation-based inference. None of the mitigation strategies they tested prevented failure in every case—a finding about that study, not a verdict on every subsequent method.
Choose evidence that could change the conclusion
Compare candidate models with observations, then look for a measurement that would separate their predictions. Murphy’s research framework connects model proposals with experiment selection. For a research team, the useful next observation is one where plausible alternatives differ detectably, taking account of measurement noise and feasibility.
Taken together, these findings suggest five questions for reviewing an AI-generated scientific explanation:
What is the claim? Distinguish reproducing existing measurements, predicting new outcomes and identifying a mechanism. Evidence for one should not silently become evidence for all three.
What alternatives were considered? Request the candidate models, the data they used and the assumptions that constrained the search. Include priors—the starting assumptions about plausible values—and reasons for excluding alternatives.
What has been checked? Ask separately for output comparisons, uncertainty diagnostics and checks of the simulator’s adequacy. A convincing plot does not answer all three questions.
What could distinguish the survivors? Have a domain expert assess a feasible observation or intervention. Record each candidate’s prediction before obtaining the result, including whether the difference is large enough to detect.
What remains unresolved? State which explanations the evidence favours and which remain plausible. If none of the candidates predicts the new observation adequately, reconsider the model assumptions rather than merely declaring a winner.
This is a different question from how often scientists use an assistant. RohitAI’s analysis of researchers’ Claude usage examines that distinction between adoption and demonstrated research progress. For scientific modeling, the immediate decision is whether a candidate warrants further testing and which comparison would be informative.
If two mechanisms remain indistinguishable, preserve both in the research record. That unresolved disagreement can help determine what to measure next.