A model can predict what people do and still leave an important question open: what will they do if we change the situation?
This is where my interest in modelling human cognition becomes quite practical. A model may be useful for predicting the next response, explaining a pattern, or choosing an intervention. Each use asks for a different kind of evidence. I am especially interested in the step from the first to the third.
The prediction that matters↗
Imagine two models of cooperation. One responds mainly to recent rewards; the other is sensitive to a longer history of perceived fairness. Under the conditions in an existing dataset, their predictions could be very similar. A new allocation rule might separate them: the same change in incentives could carry different implications for a model that tracks a social norm.
This is a hypothetical example. The point is that agreement on recorded behaviour can leave disagreement about a proposed change unresolved.
For intervention selection, the target might be the difference between expected human outcomes under two specified policies:
The target population, setting, and time horizon are held fixed. is the outcome under policy , and is its expectation in that population. compares the policies; the observed difference between people who received them is not automatically the same quantity.
A model does not necessarily need to predict every individual action accurately to get this comparison right. Conversely, a good average fit to existing actions does not establish that it ranks the two policies correctly. The relevant uncertainty is attached to the decision we want to make.
There are already concrete examples of using artificial behavioural models in this way. Koster and colleagues trained neural models of human players in an iterated multiplayer trust game, used reinforcement learning to design a common-pool allocation mechanism, and evaluated it with human participants. This is validation within a particular game; it does not establish reliability for real-world policy decisions. That makes the next question more interesting: under which conditions should such a transfer work, and how could we detect a failure before relying on it?
A comparison I would find informative↗
I would start with two models that perform similarly on a specified behavioural prediction task but make meaningfully different predictions about a held-out intervention. The intervention must have independent human outcome data capable of adjudicating that disagreement.
“Similarly” needs some care. An average score can hide differences in calibration, participant subgroups, interaction histories, or the situations each model has seen. If one model has richer information, its advantage may tell us little about the proposed explanation. I would want to make the information budget explicit, along with the capacity and fitting choices relevant to the comparison.
The decisive outcome would concern the intervention: its direction of effect, the ranking of candidate policies, or the loss from choosing the policy recommended by the model. For a finite candidate set , with higher outcomes preferred, that last question can be written as:
is the model’s estimated policy value, and is its recommendation. compares that recommendation with the best candidate policy for the actual human population. This is a proposed evaluation quantity; estimating it still requires independent human evidence and an account of uncertainty.
I would also compare against what could be learned from the same amount of human data without a synthetic model. The value of the model has to be assessed against an alternative use of the available evidence.
A disagreement alone would be an incomplete result. The useful finding would identify a condition that helps explain or anticipate which recommendation transfers—for example, the range of interaction histories on which it was validated. A condition found in one task would then need to survive another test.
What would count as understanding?↗
The intended correspondence needs to be explicit. Are we asking whether a model reproduces human choices, whether particular processes play corresponding causal roles, or whether it implements the same physical mechanism? Evidence for one of these claims does not settle all of them.
Causal abstraction offers a useful example of making a correspondence testable: Geiger and colleagues align neural representations with variables in a higher-level causal model and use interventions to examine that alignment. Their study concerns neural networks and a specified causal model; it does not by itself establish a correspondence between an LLM and a human mind.
For a human-facing decision, the required correspondence could be narrower. We might only need reliable predictions for a defined family of interventions, in a defined population and setting. That would be valuable even while larger questions about mechanism remain open.
The research question I keep returning to is therefore conditional: what additional evidence makes a model's recommendation worth acting on, beyond its fit to what people have already done? It gives the comparison between humans and machines a concrete job, and gives that comparison a way to fail.