CONCEPT
Disembodied Prediction
Andy Clark’s diagnosis of the fundamental limitation of current AI: a generative model that predicts without the grounding in embodied action and sensory consequence that keeps the brain’s predictions tethered to reality—the structural source of confident, fluent hallucination and the architectural reason the human component of a human-AI system is not merely useful but irreplaceable.
The brain and the large language model are both, at the most abstract computational level, prediction machines. Both learn to generate expectations about inputs; both update when reality deviates from expectation; both produce outputs shaped by vast bodies of training data.
Andy Clark’s
predictive processing framework illuminates what they share—and what distinguishes them in a way that is not an engineering detail but an architectural difference of kind. The brain’s predictions are biological: they are constrained by
the goal of selecting the right actions at the right times, tethered to a body whose survival depends on accurate modelling of the world, updated by the sensory consequences of the organism’s own actions. When the brain’s predictions are badly wrong, the prediction errors are large, the organism is disoriented or injured, and the model is updated with urgency. The AI’s predictions