Immersion as inference: An active inference account of induced realities and its relation to the two-competitor model
Immersive technologies are designed for a single target: the human perceptual system. Engineers of virtual and augmented reality have long recognized that a display does not have to replicate the physical world, it only has to provide the inputs that satisfy the senses at the level at which the brain samples them. This engineering intuition is, in fact, a statement about brain function. Here we develop it into a formal account of immersion within the active inference framework. We treat an induced reality (a virtual, augmented, or mixed reality, but also a book or a daydream) as a candidate generative model that competes with the model of the actual reality for the explanation of a shared sensorium and for the control of action. Immersion is then defined as the ensuant belief that the induced reality best explains the current sensory signals and supports the current perceived reality of the winning model, despite any break in immersion by Bayesian model switching driven by a high-precision prediction error that the induced model cannot absorb. Exogenous and endogenous attention, implemented as weights governing stream selection in our earlier two-competitor model, which treated attention as necessary but not sufficient for immersion, appear as two routes of precision allocation: likelihood precision of evolutionarily entrenched sensory events, and precision that follows the expected free energy of policies. Affect and motivation enter through the pragmatic and epistemic value of remaining in the induced reality and through the rate of change of free energy. We show that the two-competitor attentional focus vector and its angle are recovered from the active inference variables, and we reproduce the two scenarios of the original model, a rewarding and a fatiguing activity interrupted by a slamming door, in a minimal simulation. The account links presence, embodiment, cybersickness, and narrative absorption to a single quantity, and it yields concrete predictions for the psychophysiological assessment and closed-loop control of immersion. We discuss implications for neuroergonomic design in which the evolutionarily entrenched core of perception is non-negotiable while its experience-dependent periphery can be trained.