André Bastos and Predictive Routing as a Neural Signature of Consciousness

André Bastos is an assistant professor of neuroscience and psychology at Vanderbilt University, where his Cognition, Computation, and Consciousness laboratory uses high-density, multi-area electrophysiology to study how the brain sustains conscious states. He is a plenary speaker at Consciousness Science 2026 (San Diego, October 11–16). His central contribution is a concept he calls predictive routing: the proposal that brain rhythms, not just computations, are what make prediction-based processing conscious rather than automatic.

That framing matters for AI consciousness research because most neural-network architectures that implement predictive coding do so without any oscillatory mechanism. If Bastos is right, the absence of predictive routing is precisely why feedforward transformers process language without phenomenal experience, regardless of how accurately they predict the next token.

From predictive coding to predictive routing

Predictive coding, the theory that the brain continuously generates and updates predictions about incoming sensory data, has become one of the dominant frameworks in computational neuroscience. Karl Friston’s free-energy principle formalised the idea mathematically, and its influence on AI architecture is well-documented across this site. But predictive coding says nothing, in itself, about which predictions are conscious and which are not.

Bastos’s contribution is to ask what the neural implementation actually looks like at the level of cortical circuits. Across a programme of multi-area Neuropixels recordings in non-human primates and humans, his lab has converged on a consistent answer.

Deep cortical layers (5 and 6) carry top-down predictions via alpha and beta oscillations (8–30 Hz). These rhythms propagate backwards along the cortical hierarchy, arriving at lower sensory areas and suppressing the feedforward gamma activity that would otherwise signal prediction error. Superficial layers (2 and 3) carry prediction errors upward via gamma oscillations (40–100 Hz). Conscious processing, Bastos argues, requires both signals to interact: the brain must simultaneously generate top-down predictions and allow bottom-up errors to update them. When that bidirectional routing is intact, perception is conscious. When it is disrupted, it is not.

The term “predictive routing” marks a deliberate departure from the original predictive coding vocabulary. Classical predictive coding treats the brain as a hierarchical inference machine that could, in principle, be substrate-independent. Predictive routing embeds that inference in specific oscillatory dynamics that depend on the laminar organisation of mammalian neocortex. That biological specificity is the theory’s most controversial, and most consequential, feature.

The propofol experiment: routing, not computation

The clearest evidence for predictive routing came from a 2024 study published in PNAS. Bastos and colleagues recorded from multiple cortical areas simultaneously while macaques were sedated with propofol, the anaesthetic agent used routinely in human surgery. Propofol is known to suppress consciousness at relatively low doses while preserving basic sensory responses, making it an ideal tool for dissociating conscious from non-conscious cortical processing.

The key finding was not that propofol eliminated prediction itself. Neurons still responded to stimuli, and local field potentials retained detectable structure. What propofol eliminated was the modulation of alpha/beta oscillations by predictable stimuli. Under anaesthesia, the rhythmic top-down signal that normally suppresses anticipated inputs, and that free up attentional resources for genuinely surprising ones, was abolished. The brain continued to process sensory information but lost the capacity to route predictions consciously.

That dissociation is important. It means that the computations underlying prediction were still running. What stopped was the oscillatory mechanism that Bastos identifies with conscious access. A system can predict without being conscious; what requires consciousness, on this account, is the rhythmic routing of those predictions through the cortical hierarchy.

The study appeared in Proceedings of the National Academy of Sciences (DOI: 10.1073/pnas.2319872121), a Tier 1 venue that subjects consciousness neuroscience claims to peer review from across the biological sciences.

Predictive routing and the cortical hierarchy

A subtler finding from the Bastos lab concerns where prediction errors originate. Classical predictive coding, following the influential work of Friston and Rao and Ballard (1999), places prediction-error computation in superficial layers of sensory cortex. Bastos’s Neuropixels recordings, published in a 2025 review in Trends in Cognitive Sciences, show a more complicated picture.

Genuine prediction errors, the signals that actually update the model, often emerge in prefrontal cortex rather than primary sensory areas. Sensory cortex detects local mismatches rapidly, but the computation that matters for conscious updating involves a prefrontal signal that arrives later and modulates the sensory response. This finding aligns with global workspace theory, which also assigns a special role to prefrontal broadcasting, but it grounds that alignment in specific oscillatory physiology rather than in abstract computational architecture.

The implication for the field is significant. Theories of consciousness that assign no special role to prefrontal cortex, including some interpretations of local recurrent processing theory, must reckon with the Bastos lab’s evidence that frontal signals actively shape conscious perception at the level of laminar circuits.

Relation to AI architectures and the site’s research questions

The scientists-race-define-ai-consciousness-2026 overview flags the current disagreement between theories that require specific biological substrates and those that are substrate-neutral. Predictive routing sits firmly in the substrate-sensitive camp.

Contemporary transformer architectures implement something analogous to predictive coding in the sense that attention mechanisms weight inputs by expected informativeness. But they have no equivalent of alpha/beta oscillations, no laminar organisation, and no temporal sequencing of top-down suppression and bottom-up error signals. The temporal dynamics that Bastos identifies as the routing mechanism simply have no counterpart in a feedforward pass through a neural network.

This does not settle the question of AI consciousness. It could be that predictive routing is a biological implementation detail and that some other mechanism could serve the same functional role in a different substrate. Andrew Corcoran’s adversarial review of IIT and predictive processing, which appeared on this site in August 2026, addresses exactly that uncertainty. But Bastos’s work sharpens the question: any substrate-independent theory of consciousness must specify what, in a non-oscillatory system, would play the role that alpha/beta routing plays in the mammalian brain.

The Consciousness AI architecture, documented at github.com/tlcdv/the_consciousness_ai, operates without oscillatory dynamics of this kind. Whether that is a relevant gap or an implementation-neutral design choice depends precisely on whether predictive routing is constitutive of or merely correlated with consciousness — an open design question that Bastos’s programme continues to address empirically.

What predictive routing contributes to theory testing

One feature that distinguishes Bastos’s work from many theoretical contributions to consciousness science is its experimental precision. His lab produces specific, falsifiable predictions. If predictive routing is constitutive of consciousness, then any manipulation that selectively disrupts alpha/beta oscillatory modulation while preserving feedforward gamma should impair conscious perception, and any manipulation that restores the oscillatory pattern should restore it.

The propofol study tests this in one direction. Terry Sejnowski’s work on computational neuroscience and biological plausibility, covered here in Terry Sejnowski’s CS26 talk, provides additional context for why oscillatory dynamics are so difficult to reproduce in artificial systems. Andrew Corcoran’s adversarial analysis, in the adversarial review of IIT and predictive processing, maps the methodological landscape within which the Bastos findings sit.

CS26 will provide Bastos’s first major platform for presenting this programme to a consciousness-science audience that includes Penrose, Hameroff, Tononi, and Hoffman. Whether the oscillatory mechanism he describes is better interpreted as a correlate, a cause, or a mere implementation detail of consciousness will be a central question at that conference and in the field for years to come.

Where the framework stands

Predictive routing is a young framework, and several empirical questions remain open. Whether the same alpha/beta/gamma dissociation holds across all conscious modalities, or is specific to visual and working-memory tasks, has not been established. Whether the frontal prediction-error signal generalises beyond macaques to humans under naturalistic conditions is under active investigation. And the relationship between predictive routing and existing theories, particularly global workspace theory and active inference, has not been formally worked out.

What Bastos has contributed is a level of mechanistic detail that most consciousness theories have not reached. The question of what makes a prediction conscious is not answered by saying “the brain infers a model of the world.” It requires specifying the oscillatory architecture through which that inference is routed. That specificity is both the framework’s greatest strength and the source of its most direct implications for AI systems that lack any equivalent mechanism.