From root: Why should I simulate the real world?

Why keep humans in the loop with world models?

To bring in knowledge that the data alone cannot supply. In order to tackle data that is irrelevant, incomplete, scarce or wrong, it helps to bring as much useful information into the problem as possible. This information could come in the form of inductive biases in the model structure or prior knowledge provided by an expert.

Black box AI at best encodes only generic statistical or architectural inductive biases, but usually lacks the rich, domain‑specific structure that an explicit ‘world model’ can provide.

In contrast to black box AI, world models represent the state of the system separately from the data. Simulations can be thought of as a kind of world model, for instance.

A true world model should enable sources of information distinct from the data to be included in its prediction, i.e., domain-specific structure with expert interpretation. This is the only clear way to circumvent data limitations.

Humans ‘in the loop’ can also interrogate their world models to find explanations for their predictions. In this sense, world models can provide insights, not just predictive outputs.

References and further reading

Ha, David, and Jürgen Schmidhuber. 2018. World Models. arXiv preprint. https://arxiv.org/abs/1803.10122.

Hardwick, Robert J. 2025. “What can a simulation do that a prediction cannot?” umbralcalculations, November. https://umbralcalc.github.io/posts/what_can_a_simulation_do_that_a_prediction_cannot.html.

Hardwick, Robert J. 2026. “When is data good enough for black box AI?” umbralcalculations, November. https://umbralcalc.github.io/posts/when_is_data_good_enough_for_black_box_ai.html.