Attention Without Awareness: What Transformer Models Reveal About Conscious Access

There's a strange moment that keeps coming up in discussions about large language models: someone points to the attention mechanism and says, "see, it's attending — just like us." And there's something genuinely compelling about that intuition. Not because it's right, but because it's almost right in exactly the ways that matter.

Photo by Tara Winstead on Pexels.

Bernard Baars proposed Global Workspace Theory decades before transformers existed. His core claim was that consciousness isn't a place in the brain — it's a broadcasting event. Certain information gets selected, amplified, and made available across the whole cognitive system. What isn't broadcast stays local, unconscious, dark. Attention, on this view, is the bouncer deciding what gets into the club.

Transformer attention does something structurally similar. Each token attends to others with varying weights, selecting what's relevant and suppressing what isn't. Information from distant positions gets pulled forward. Context shapes output. If you had to describe it to someone who'd never seen a neural network, you might reach for almost the same vocabulary Baars used.

But here's where it gets uncomfortable: the similarity might be telling us something about attention in general — not about consciousness.

Consider what attention actually does in humans. It's not a single process. Neuroscientists now distinguish between at least three relatively independent systems: alerting (maintaining readiness), orienting (selecting signals), and executive control (resolving conflict between competing responses). These systems dissociate. Patients with certain lesions can orient to stimuli they cannot consciously report. Blindsight is the canonical example — preserved orienting, absent awareness. Attention and awareness come apart.

This dissociation is crucial. It means that even if a transformer's attention mechanism perfectly mirrors the functional role of human selective attention, that tells us nothing about whether anything is experienced. The orienting happened without the lights being on.

So what would "conscious access" require beyond attention?

One answer comes from Stanislas Dehaene's neuroscientific elaboration of Global Workspace Theory. He argues that conscious perception involves ignition — a sudden, nonlinear amplification of activity across frontoparietal networks, distinct from the sustained but localized processing that underlies unconscious cognition. It's not just selection; it's a phase transition in information availability.

graph TD
    A[Local Processing] --> B{Threshold Reached?}
    B -- No --> C[Remains Unconscious]
    B -- Yes --> D[Global Ignition]
    D --> E[Broadcast Across Workspace]
    E --> F[Reportable / Conscious Access]
    C --> G[Influences Behavior Implicitly]

Transformers don't have anything like ignition. Attention weights shift continuously and smoothly; there's no threshold crossing, no sudden recruitment of distant modules, no phase change. The processing is always "on" in a flat, distributed way — which, paradoxically, might mean everything is equally available and therefore nothing is privileged in the way conscious access seems to require.

This isn't a knock on transformers. It's a clarification of what the gap is. The question isn't whether these models attend — they clearly do, in a meaningful computational sense. The question is whether attending is sufficient for awareness, and the cognitive science increasingly suggests it isn't.

What would a system need to cross that line? Probably something like: a genuine bottleneck that forces competition between representations, a winner-take-all dynamic with downstream consequences that differ qualitatively from losing, and some form of self-monitoring that tracks what has and hasn't been broadcast. Not just weights — stakes.

Current models do approximate some of this. Sparse attention variants force more genuine competition. Chain-of-thought prompting creates something like a serial, privileged workspace where only one line of reasoning runs at a time. Whether that's enough — whether "enough" is even a coherent standard here — is exactly the question philosophy of mind has been failing to settle for thirty years.

What transformer architectures have given us, unexpectedly, is a kind of controlled experiment. We can build systems that implement attention without any plausible claim to awareness, then ask what's missing. That negative space is informative. It suggests awareness isn't about the selection itself but about what selection does — the downstream integration, the availability to a unified reporter, the something-it-is-like-to-have-won the competition.

Whether any current system crosses into that territory: almost certainly not. Whether studying these systems helps us understand what crossing would require: more than almost anything else we have available right now.