Can predictive processing be integrated with global neuronal workspace theory into a single working hypothesis for conscious access?

Why would anyone combine these two theories?

Predictive processing treats perception as a guessing game: the brain constantly generates predictions about what is out there and compares them with incoming sensory signals. The difference between prediction and input is an error signal, which the brain uses to update its guesses. Global neuronal workspace theory, by contrast, is about access: it proposes that information becomes conscious when it is broadcast widely across a network of brain regions, especially prefrontal and parietal areas, in a sudden 'ignition' event. The two theories answer different questions—one about the content of perception, the other about when that content becomes consciously available—so combining them is not redundant.

The integration is not just a philosophical exercise. In schizophrenia, patients can show impaired conscious access alongside intact unconscious processing. One account within the workspace framework attributes this to a disruption in top-down attentional amplification, and confronting that account with Bayesian models led researchers to encourage combining the two theories [1]. The proposed integration treats the workspace as the stage where predictive models become conscious, and treats predictive coding as the machinery that generates the models being broadcast.

What does the combined hypothesis actually look like?

In the sketch offered by Alegiani and Marraffa, predictive coding provides a hierarchical generative model of the world, and the global workspace provides a mechanism for selecting and broadcasting one of those predictions at a time [1]. Conscious access, on this view, occurs when a prediction (or a prediction error) is amplified strongly enough to ignite the workspace and become available to many brain systems. The authors are explicit that the relation between conscious access and Bayesian hierarchical modeling still needs to be made clear, so the integration is a direction of travel rather than a finished model [1].

This combined hypothesis makes a concrete prediction: if conscious access depends on broadcasting a prediction, then disrupting the top-down amplification that carries predictions should impair conscious access while leaving unconscious processing intact. That is exactly the pattern the schizophrenia account describes, which is why the clinical case is used as a testing ground for the integration [1]. The authors also note that delusions can be read as a failure of predictive updating within this framework, though they present this as a direction for further work rather than a settled explanation [1].

Does the neural evidence support the workspace half of the hypothesis?

The workspace half has recently received direct support from a re-analysis of magnetoencephalography data from 100 participants viewing faces. Using dynamic causal modeling—a method that estimates directed influences between brain regions rather than mere correlation—researchers found evidence for the workspace theory's predicted prefrontal 'ignition' at both the onset and the offset of a stimulus [2]. That is important because the offset ignition had not been seen in the original analyses, so the workspace prediction held up under a more mechanistic test.

However, that study tested the workspace theory against integrated information theory, not against predictive processing, and it did not incorporate predictive coding into its framework [2]. So it strengthens the case that a workspace-like broadcast happens, but it does not tell us whether the content being broadcast is a prediction. The same study also found only weak evidence for face selectivity in its networks, which the authors say warrants a conservative interpretation of the competing theory's findings [2]. For the combined hypothesis, the takeaway is that the workspace mechanism looks real, but its link to predictive content remains untested in this dataset.

What would it take to turn the sketch into a working hypothesis?

The integration faces at least two open challenges. First, the relationship between conscious access and Bayesian hierarchical modeling has not been made explicit, which the proponents of the integration themselves flag as a critical gap [1]. Second, the workspace theory is itself being refined: a recent multilevel account emphasizes that global neuronal workspace theory extends from cellular and molecular mechanisms to large-scale network dynamics, and argues that it is not a functionalist computational theory in the way the older global workspace idea was [3]. If the workspace is more than a computational broadcast, then plugging predictive coding into it may require more than treating the workspace as a simple stage.

A separate proposal, emergent flow theory, treats both global neuronal workspace theory and predictive processing as existing frameworks but does not integrate them; instead it adds autonomic coherence as a gain parameter that modulates transitions between temporal regimes of conscious integration [4]. That work is a reminder that other researchers see the gap differently—not as a missing link between prediction and broadcast, but as a missing physiological control parameter. Meanwhile, the large adversarial collaboration between workspace and integrated information theory has been designed to test those two theories against each other, not to test predictive processing [5]. So the combined hypothesis remains a promising sketch with clinical motivation, but it has not yet been put to a direct neural test.

Sources used in this answer

Related research questions