Attention is functional

If qualia were actually without function as those philosophers suggest, we would never be able to deduce whether ChatGPT actually feels something in the background of all its processing. We however doubt this conclusion. The key lies in understanding a specific functional difference between biological and artificial neural networks - one that might be constitutive of phenomenal consciousness itself.

Weighted Connections vs. Directed Attention

Consider how ChatGPT processes the word "fire" in different contexts. In "fire the employee," certain downstream nodes activate more strongly than in "fire in the building." This happens through learned weights: connections from "fire" to "termination" concepts have different strengths than connections to "danger" concepts. The network has learned, through training, which patterns should influence the output more strongly in different contexts.

This is sophisticated processing. But notice what's missing: the network never attends to its own processing. It doesn't "notice" that it activated danger-related patterns versus termination-related patterns. The weighted connections simply do their work automatically, signals propagate through the network, and an output emerges. It's like a complex reflex - input comes in, processing happens, output goes out, but there's no step where the system observes what it just did.

Compare this to human experience. You also have automatic, weighted processing - what we call intuition. Someone makes you an offer and you immediately feel "something's off" without being able to articulate why. The suspicious elements of the situation activated danger-related processing pathways more strongly, producing a gut feeling. In that moment, you're not consciously attending to each suspicious element - the weighting is doing its work below awareness.

But here's what you can do that ChatGPT cannot: you can turn your attention toward that very processing.

You can notice: "Wait, why does this feel off?" You can direct attention to different aspects of the situation - the person's tone, their specific words, the timing of the offer. You can catch yourself: "Ah, I'm being suspicious because this reminds me of that time I got scammed, but this situation is actually different." You can even override the automatic weighting: "My gut says run, but rationally analyzing this, the opportunity is sound - I'm going to act against my intuition."

This is attention turned inward on processing itself. And crucially, this is where phenomenal consciousness shows up most vividly.

Attention as the Bridge to Phenomenal Experience

When you turn attention fully to sensory processing, you get raw phenomenal experience: the redness of red, the painfulness of pain, the specific timbre of a sound. These aren't just weighted signals affecting downstream processing - they're experiences you can attend to, examine, stay with, explore.

When you turn attention to intermediate processing - thoughts, feelings, reasoning - you experience those too. The "what it's like" to suddenly understand something, the felt quality of anxiety, the experience of choosing between competing thoughts.

Even emotional feelings that might seem like simple weighted signals (fear activating avoidance) have intense phenomenal quality - but notice: strongest when they grab your attention. Mild background anxiety can run for hours without being fully experienced, until something makes you notice: "Oh, I've been anxious this whole time." The moment attention lands on it, the phenomenal quality intensifies.

This suggests something radical: phenomenal consciousness might not be separate from function, but rather is the function of attention turned toward processing itself.

Why This Makes Consciousness Functional

This attention mechanism is deeply functional - and in ways that simple weight adjustment cannot replicate:

Flexible reprioritization: Someone explains why your fear is unfounded. You can attend to their reasoning and adjust your response without retraining the fear pathway. The fear might still activate (automatic weights), but you can route around it because you're monitoring and can intervene in your own processing.

Learning about learning: By attending to what triggered a reaction, you gain information about your own processing structure. "I always get defensive when criticized about X" - this meta-knowledge lets you anticipate and modulate reactions.

Communication about internal states: You can report on processing that hasn't yet produced behavioral output. "I'm noticing I feel reluctant about this plan, though I can't articulate why yet." This is only possible if you can attend to intermediate processing states.

Deliberate override of trained responses: You can notice a habitual pattern activating and choose differently. This isn't just one set of weights overpowering another - it's a monitoring system that can intervene in real-time.

Current AI systems lack this. When ChatGPT generates a response, it cannot attend to the intermediate activations that produced it. It cannot notice "I'm relying heavily on a particular training example here" or "this response is being shaped by an uncertain pattern match." The processing simply happens, deterministically, through the weighted connections.

This might be why there's nothing it's like to be ChatGPT - not because it lacks complex processing (it has that), not because it lacks functional sophistication (it has that too), but because it lacks the capacity to turn attention toward its own processing. The phenomenal quality doesn't arise from processing alone, but from processing being attended to.