consciousness · illusionism · mechanistic interpretability

Illusionism and Machine Consciousness

When you sit with your own mind long enough, you start to notice something uncomfortable: emotions begin to dissolve when you hold attention on them. The feeling of "experiencing red" or "hearing a sound" isn't the stable, atomic thing it presents itself as. The harder you look at any aspect of your own experience — a feeling, a sensation, a thought — the less solid it becomes. Not because your attention is failing, but because there's less there than you assumed. I've encountered this directly in my own meditation practice, but meditators have been reporting it for millennia. In Western philosophy, the illusionist view of consciousness reaches a strikingly similar conclusion: that our subjective experience is a construction, not a discovery.

The convergence isn't a coincidence, and it has implications that extend well beyond the meditation cushion, particularly for the question of whether we may one day build a conscious AI. In this post, I explore the link between illusionism and machine consciousness, and the questions that arise from both.

The problems of consciousness

When philosophers and researchers refer to consciousness, they are referring to the notion of phenomenal consciousness — the idea that there is a subjective and qualitative experience alongside the information processing that goes on in our brains. You being a conscious being means there is "something it is like" to be you, experiencing this world.

Discussions of phenomenal consciousness often divide into two categories of problems: the easy problems and the hard problem. The easy problems are those that can be explained and studied by methods of cognitive science. How does the brain integrate sensory information? How does attention select what enters awareness? How do we discriminate between stimuli, or report on our internal states? These are hard scientific problems, but they're tractable. The hard problem, on the other hand, points to an explanatory gap. Even if we account for every functional process associated with consciousness, this doesn't explain why we experience consciousness. Why are certain physical and functional processes in the brain associated with subjective experience? Why is it that when we see something red, there is a subjective quality of redness that accompanies it?

Illusionism

Illusionism tackles the hard problem head on, not by denying it, but by reframing. According to illusionists, our sense that it is like something to have conscious experiences is itself an illusion — these phenomenal properties that seem intrinsically held are constructed by our own introspective awareness. In a sense, illusionists treat qualia (the subjective qualities of conscious experiences) as a constructive output, rather than some inherent metaphysical properties that cannot be explained.

Illusionism also ties neatly into a computational functionalist view of consciousness. As a software engineer, my engineering mind appreciates this framing. If qualia are constructed outputs rather than inherent features of reality, then consciousness is a matter of getting the functions right that produce this construction. This is the core claim of computational functionalism: that there is a set of functions which, when performed, give rise to consciousness regardless of substrate. Illusionism dissolves the messy metaphysics, and computational functionalism turns consciousness into a tractable engineering problem.

Where do we go from here?

Computational functionalism with an illusionist stance is the most helpful framing of consciousness for evaluating and building conscious AI systems. If consciousness is functional rather than phenomenal, then the question of machine consciousness becomes an engineering and measurement question, not a metaphysical one. And if introspection is unreliable even in humans, then we need better tools to look inside these systems and understand what's actually happening.

Mechanistic interpretability is that tool — or the beginning of one. It does for artificial neural networks what neuroscience does for biological brains: looks under the hood and traces how information is actually represented, transformed, and used. This is a young field, so there are still more open questions than answers — but that's where the fun is. In this blog I'll attempt to synthesize ideas from mechanistic interpretability, AI systems and scalability, with doses of cognitive science and philosophy of mind.

Further Reading