A world model can predict future observations without yet understanding which differences matter for action. Recent audits show that prediction accuracy, exploration gains, and planning success can separate. Read through situated-action and enactive theory, this is not a surprising failure of scale; it is evidence that action is constituted through ongoing coupling with a particular environment, body, and concern.

Scope: a conceptual analysis grounded in two audits

This analysis connects two preprints released 27–30 September 2026 with Lucy Suchman’s account of situated action and Ezequiel Di Paolo, Thomas Buhrmann, and Xabier Barandiaran’s enactive account of sensorimotor agency. The philosophical connection is interpretive, not a claim made by the machine-learning authors.

Accurate prediction and useful training can come apart

Does Learning to Predict the World Help Agents Act? trains agents with deliberately mismatched next-observation targets. Prediction accuracy drops sharply, yet substantial task gains remain. In tested settings, trained agents consider more actions and loop less; even random reward signals increase coverage.

The experiment does not show that environmental prediction is irrelevant. It shows that an optimisation procedure attributed to prediction may change exploration and search in other ways. A model can become a better actor for reasons its explanatory label does not capture.

The Planning Limits of Latent World Models identifies a different separation. Action-conditioned predictors rank actions reliably only within a limited horizon. Larger predictors and longer-rollout training do not simply remove the limit. Near expert subgoals help much more than unconstrained distant imagination in the authors’ manipulation tasks.

Suchman: a plan is a resource, not the action itself

In Human–Machine Reconfigurations, Lucy Suchman argues that plans organise and account for activity, but situated circumstances remain crucial to what action becomes. A plan does not contain every contingency of its execution. People use it as one resource while responding to materials, others, interruptions, and emerging consequences.

A latent rollout functions similarly. It can orient an agent toward a possibility, but it does not abolish the need to re-enter the actual situation. Model-predictive control—act briefly, observe again, replan—works precisely because imagined continuity is periodically corrected by world involvement.

Enactivism: relevance emerges in a sensorimotor relation

Sensorimotor Life develops agency through networks of sensorimotor schemes and organism–environment coupling. On this account, perception is not an internal copy of a ready-made world. Meaning depends on what an agent can do, what sustains or threatens its organisation, and how activity changes the available situation.

This clarifies why a visually faithful future may preserve the wrong distinctions. A prediction can render texture, background, and object identity while smoothing the contact transition that determines whether a grasp succeeds. Conversely, a sparse representation of force, affordance, or near-term change may look incomplete but remain action-sufficient.

Somatic practice makes the same point at human scale. A movement’s relevance is not exhausted by its visible trajectory. Balance, effort, pain, intention, and relation shape which continuation is possible. The Lab’s earlier argument that the body is not a point cloud addressed what geometric representation omits; the present issue is what prediction omits about situated consequence.

The counterargument: internal models can still support coupling

Enactive theory is sometimes used to reject representation altogether. That conclusion is not required here. Short-horizon predictions can be components inside a coupled control loop. The planning-limit study reports that selecting among eight actions proposed by a vision-language-action policy raises success from 65% to 77% across 16 tasks. A bounded model can be useful when its horizon and role are explicit.

The stronger critique concerns substitution: treating an internal rollout as if it were the world, or treating benchmark success as if it established general understanding. Situated feedback is not merely a patch for an imperfect model. It is part of the process by which action acquires meaning and remains answerable to consequences.

A different standard for world models

Instead of asking only whether a model predicts the next state, ask four questions. Which differences does the representation preserve? Over what horizon do those differences remain decision-relevant? What observation interrupts a wrong imagination? Whose goals determine that an outcome counts as success?

These questions move evaluation from resemblance toward accountability. They also connect to this month’s frontier report on action-relative representations: the best world model may not be the one that imagines furthest, but the one that knows when its imagined distinction has stopped being useful.

References