Computational modeling has become a defining feature of contemporary psychological science. From Bayesian accounts of perception to reinforcement learning models of decision-making, formal frameworks now permeate domains once dominated by verbal theorizing. Yet the philosophical question of what these models actually explain remains surprisingly underexamined.

There is a temptation to equate mathematical precision with theoretical depth. A model that fits behavioral data with impressive accuracy can feel like it must be capturing something real about the mind. But fit is not understanding, and prediction is not explanation. A model can mimic phenomena it does not illuminate, just as a planetarium reproduces stellar motion without revealing celestial mechanics.

The question becomes urgent as computational approaches consolidate institutional authority within psychology. If we cannot articulate what makes a model genuinely explanatory, we risk mistaking formal sophistication for theoretical progress. We may build elaborate apparatus around phenomena while remaining as puzzled about underlying processes as before. The discipline of distinguishing what models do—and what they merely appear to do—is therefore not pedantic philosophy but methodological hygiene essential to cumulative knowledge.

Computational models in psychology serve at least four distinct epistemic functions, and conflating them generates persistent confusion. Prediction concerns whether a model accurately forecasts observable outcomes. Simulation involves reproducing patterns of behavior or neural activity. Exploration uses models as conceptual sandboxes for examining the implications of theoretical assumptions. Explanation, the most contested function, claims to illuminate why phenomena occur as they do.

These functions are logically independent. A model can predict superbly without explaining anything, as countless atheoretical regression models demonstrate. Conversely, a model may offer genuine explanatory insight while predicting poorly, particularly when explaining mechanisms whose effects are obscured by noise or contextual variation.

Simulation occupies a curious middle ground. A model that successfully simulates a phenomenon shows that some mechanism can produce the observed pattern, but not that the actual mechanism resembles the modeled one. The phrase sufficiency does not imply necessity captures this constraint. Many architectures can generate similar surface behavior.

Exploratory modeling, perhaps the most epistemically honest function, treats models as instruments for thinking. By formalizing assumptions, researchers discover hidden implications, internal inconsistencies, or unexpected consequences of their theoretical commitments. The model's value lies in what it teaches the modeler, not in any claim about psychological reality.

Recognizing this plurality matters because researchers often slide between functions without acknowledgment, presenting predictive success as explanatory insight or treating simulation fidelity as confirmation of mechanism. Each function has its own standards of evaluation, and a model excellent at one task may be inadequate at another.

Takeaway

A model that fits the data answers a different question than a model that explains the data. Predictive accuracy and explanatory power are independent virtues, and confusing them inflates our sense of what we understand.

An explanatory model does more than reproduce data—it identifies the dependencies that make the phenomenon what it is. Following the mechanistic tradition in philosophy of science, a genuinely explanatory computational model maps its components and operations onto entities and activities in the actual cognitive or neural system. The mapping need not be exhaustive, but it must be substantive.

This requirement distinguishes how-actually models from how-possibly models. The latter demonstrate that a mechanism could in principle generate the phenomenon; the former argue that this mechanism does generate it in the system under study. Most published computational models in psychology, on careful examination, are how-possibly models presented with the rhetorical force of how-actually claims.

Explanatory adequacy also requires that the model's parameters and structural assumptions correspond to identifiable psychological or neural variables. A free parameter that improves fit but lacks any interpretive anchor is a curve-fitting device, not an explanatory posit. The proliferation of such parameters, often justified post hoc, marks the boundary where modeling shades into elaborate description.

Generalization across contexts provides another diagnostic. An explanatory model should account for phenomena beyond those used in its construction, predicting novel patterns under conditions the modeler did not anticipate. When models require extensive reparameterization for each new dataset, their explanatory pretensions weaken considerably.

Finally, genuinely explanatory models invite intervention. They specify which components, if disrupted, would alter outcomes in particular ways. This counterfactual sensitivity—the capacity to support claims about what would happen under different conditions—remains the philosophical hallmark of explanation since Woodward's influential analysis.

Takeaway

Explanation requires correspondence between model and world, not merely between model output and data. The deeper question is whether the architecture you wrote down resembles, in its joints, the system you study.

The relationship between computational modeling and verbal theorizing is often framed as competitive, with formal models cast as the mature successors to imprecise prose. This framing obscures more than it reveals. Verbal theory and computational modeling perform different epistemic labor, and psychology suffers when either is sacrificed to the other.

Verbal theories excel at articulating the conceptual landscape—identifying what phenomena are, why they matter, and how they connect to broader frameworks of human meaning. They are the medium in which problems become intelligible and worth pursuing. No computational model can be formulated until verbal theorizing has identified what is to be modeled and why.

Computational models, in turn, discipline verbal theory by exposing its hidden commitments. The act of formalization frequently reveals that an apparently coherent verbal account is ambiguous, incomplete, or internally inconsistent. This diagnostic function alone justifies modeling even when the resulting formalisms make no explanatory claims.

The deepest integration occurs when modeling and verbal theorizing operate in genuine dialogue. The verbal theory specifies the explanatory target and the conceptual constraints; the model tests whether the proposed mechanisms could produce the target phenomena; results feed back into refining the verbal account. Neither register dominates; each corrects the other.

What we should resist is the assumption that formalization automatically constitutes theoretical progress. A field can accumulate models indefinitely while remaining theoretically stagnant if those models are not connected to substantive conceptual claims. Conversely, profound theoretical advances are possible without formalization, as the history of psychology amply demonstrates.

Takeaway

Verbal theory asks what is worth explaining; computational modeling asks whether a proposed explanation could work. Each is impoverished without the other.

The computational turn in psychology has been genuinely productive, but its philosophical foundations deserve more attention than they typically receive. Models do many things, only some of which constitute explanation, and the failure to distinguish these functions has inflated our collective sense of theoretical progress.

Treating models as epistemic tools rather than theoretical destinations restores a healthier relationship between formalism and understanding. The question is not whether to model, but what we expect our models to accomplish and how we evaluate whether they have done so.

Psychology's deepest questions concern the nature of mind and meaning—questions that resist reduction to any single methodological register. Computational modeling has earned its place in the discipline's toolkit. Whether it earns the explanatory authority increasingly attributed to it depends on choices researchers make about clarity, correspondence, and conceptual integration. The mathematics will not make those choices for us.