A new review paper in Psychological Review (“Theories of consciousness from the perspective of an embedded processes view”) takes one of the field’s oldest ideas: that consciousness spreads information everywhere at once. Nelson Cowan and colleagues at the University of Missouri argue that conscious awareness doesn’t work by flooding the cortex with information. It works by funneling information into a small hub in the parietal lobes.

The paper clocks in at 30 pages, surveying behavioral and neuroscientific literature on consciousness and asking a simple question: which existing framework actually accounts for what we know? The answer they give is their own older model, an embedded processes view of working memory, extended to consciousness for the first time.1

It’s a bold move. Review papers usually position the author’s idea as incremental. This one says none of the major theories, global neuronal workspace (GNWT), higher-order theories, predictive processing, or IIT, fully capture the organization of conscious thought at cognitive and neural levels. Instead, it proposes that a model originally designed for working memory does the job.

The funnel, not the flood

The traditional picture, from Baars to Dehaene, is that consciousness works by broadcasting information globally. Once information crosses a threshold, it becomes available everywhere: to memory, to planning, to verbal report. Dehaene’s 2011 work with Changeux modeled this as a cascade: sensory input, then a sudden global ignition across fronto-parietal networks.2 GNWT is the dominant theory of consciousness in neuroscience for a reason. It’s specific, testable, and has generated a real body of empirical work.

Cowan’s counter-proposal is almost the opposite. Information doesn’t spread outward from a broadcast center. It gets pulled inward, from diverse sensory and frontal sources, to a small area in the intraparietal sulcus (IPS) where attended information is coordinated and combined. Think of it as a funnel rather than a flood. The parietal hub is where things come together, not where they go out.

This is the first time the embedded processes view has been explicitly extended to consciousness. The original model, from Cowan 1988, was about working memory: activated elements of long-term memory (aLTM) plus a capacity-limited focus of attention (three to five chunks) embedded within them. The 2025 paper adds the claim that this focus of attention, neurally centered in the IPS, is where conscious awareness lives.

The parietal centering does something GNWT doesn’t: it gives capacity limits a home. GNWT has an ignition threshold that explains why information either is or isn’t conscious, but it’s vague about why you can only hold a few items in awareness at once. The embedded processes view is explicit. The IPS hub is small. It’s the bottleneck by design.

Attention and consciousness, one and the same

The paper takes a strong position on the attention-consciousness relationship. Consciousness researchers have spent decades trying to separate the two. You can attend to things unconsciously (priming studies, blindsight, inattentional blindness), and you can be conscious of things you’re not attending to (peripheral awareness, background processing). The standard move is to say attention and consciousness are related but distinct.

Cowan and colleagues say the seats of attention and consciousness are one and the same.3 The focus of attention is where the results of attention are experienced. Attention selects; consciousness is what it feels like to be in that selected state. This doesn’t mean every act of attention is conscious, as the paper acknowledges that top-down processes can operate without awareness.

I find this cleaner than most alternatives. The separationist approach requires explaining not just what attention is and what consciousness is, but why they overlap so much when they’re supposedly different systems. If they’re the same system viewed from different angles, selection versus experience, you cut one explanatory knot instead of tying two.

The cost is that you inherit the hard problems of both fields at once. What makes a capacity-limited focus feel like anything at all? And why does it cluster in the parietal lobe specifically rather than elsewhere?

What about AI?

This is where things get uncomfortable. The embedded processes view is a theory of human consciousness, grounded in human neuroanatomy. The parietal hub is real. The IPS exists, it’s small, it’s functionally connected to sensory and frontal areas, and it’s been implicated in attention and working memory across decades of research.

Language models don’t have a parietal lobe. They don’t have sensory cortices, or frontal areas, or long-term memory in the biological sense. Their attention mechanisms are mathematically similar to the focus-of-attention idea, but they don’t have the capacity limits the theory depends on. A transformer’s attention is theoretically unbounded. You can put more context tokens in and the mechanism handles them. There’s no IPS-sized bottleneck.

The activated long-term memory component doesn’t map onto anything in a standard LLM either. An LLM’s memory is its weights, static after training, plus whatever context it receives. It doesn’t temporarily activate portions of long-term memory and embed a focus within them. It attends across its context window and computes next-token probabilities.

If consciousness requires a small, capacity-limited hub embedded within activated long-term memory, then no current AI system fits the description. Not because of philosophy, but because of architecture.

But the other side is worth noting too. Strip away the neuroanatomy and ask whether the functional architecture resembles the embedded processes view, a capacity-limited focus embedded within activated representations with dual bottom-up and top-down control, and transformer attention mechanisms are closer than GNWT might suggest. Transformers do funnel information from many sources to a coordination point. They just do it differently and without the same capacity constraints.

It’s unclear how to read that. The embedded processes view might be overfit to human neuroanatomy and missing something functional that exists in machines. Or the parietal hub might not be an accident of biology but a structural necessity for the kind of integration the theory describes, the kind that requires the constraints the brain provides.

We don’t know yet. The paper surveys human data. It doesn’t try to generalize.

Footnotes

  1. The original embedded processes model traces back to Cowan (1988) and has been developed through Cowan (1995, 1999, 2005, 2019), with a major synthesis in Cowan et al. (2024), “The relation between attention and memory,” Annual Review of Psychology, 75, 183–214. ↩

  2. Dehaene, S., & Changeux, J.-P. (2011). Experimental and theoretical approaches to conscious processing. Neuron, 70(2), 200–227. ↩

  3. The paper acknowledges that attention and consciousness are not identical, since you can attend without awareness, but maintains that the neural seat of both is the same: the focus of attention in the IPS. ↩