The Mind–Body Problem and What AI Forces Us to Rethink

Descartes famously divided reality into two kinds of substance: res cogitans (thinking substance, mind) and res extensa (extended substance, matter). The mind was immaterial, unified, self-transparent. The body was mechanism.

This dualism has largely been abandoned by contemporary philosophy — but the problems it was gesturing at haven't gone away. In fact, AI sharpens them considerably.

The Hard Problem#

David Chalmers distinguishes the "easy problems" of consciousness — explaining how the brain integrates information, discriminates stimuli, focuses attention — from the hard problem: why is there something it is like to be a creature with those functional capacities?

Even a complete functional account of how visual cortex processes color leaves open the question: why does experiencing red feel like anything at all?

"The really hard problem of consciousness is the problem of experience. When we think and perceive, there is a whirl of information-processing, but there is also a subjective aspect." — David Chalmers, The Conscious Mind (1996)

This is not a gap that more neuroscience straightforwardly fills. It's a conceptual gap: the explanatory tools we use for physical processes (causal roles, functional relations) seem not to touch the intrinsic, qualitative character of experience.

Functionalism and Its Discontents#

The dominant view in philosophy of mind today is some form of functionalism: mental states are defined by their causal/functional roles, not by what they're made of. Pain = the state caused by tissue damage, causing avoidance behavior, etc. Silicon or neurons, doesn't matter — it's the pattern, not the substrate.

If functionalism is right, AI systems in principle could be conscious — they just need the right functional organization.

But Ned Block famously distinguished:

• Access consciousness: information being globally available for reasoning, report, and control
• Phenomenal consciousness: the "what it's like" aspect, qualia

A system can be access-conscious without being phenomenally conscious. This is what Block's China Brain thought experiment targets: 1.4 billion people each simulating one neuron in a network — does the whole system have inner experience?

Integrated Information Theory#

Giulio Tononi's Integrated Information Theory (IIT) takes phenomenal consciousness seriously as a fundamental feature of the world and tries to make it measurable.

The core claim: consciousness = integrated information, $\Phi$ (phi). A system is conscious to the degree that it is more than the sum of its parts — the information generated by the system as a whole, above what's generated by its parts independently.

import numpy as np
import itertools

def phi_approximation(tpm):
    """
    Rough illustration of IIT phi calculation.
    tpm: transition probability matrix, shape (2^n, 2^n)
    Returns a rough proxy for integration.
    """
    n = int(np.log2(tpm.shape[0]))
    
    # Full system entropy
    marginal = tpm.mean(axis=0)
    marginal = np.clip(marginal, 1e-12, None)
    h_full = -np.sum(marginal * np.log2(marginal))
    
    # Average over bipartitions
    partition_entropies = []
    for k in range(1, n):
        part_a = list(range(k))
        part_b = list(range(k, n))
        # Marginalize (very rough proxy)
        size_a = 2 ** len(part_a)
        size_b = 2 ** len(part_b)
        h_ab = 0
        for sub in [size_a, size_b]:
            p = np.ones(sub) / sub
            h_ab += -np.sum(p * np.log2(p))
        partition_entropies.append(h_ab)
    
    min_partition_entropy = min(partition_entropies)
    phi_proxy = h_full - min_partition_entropy
    return phi_proxy, h_full

# Simple feedforward-like vs. recurrent-like TPMs
np.random.seed(1)
n_states = 8

# Feedforward: rows roughly independent
ff_tpm = np.random.dirichlet(np.ones(n_states) * 0.5, size=n_states)
phi_ff, h_ff = phi_approximation(ff_tpm)

# Recurrent: more correlated
rec_base = np.eye(n_states) * 3 + np.random.rand(n_states, n_states)
rec_tpm = rec_base / rec_base.sum(axis=1, keepdims=True)
phi_rec, h_rec = phi_approximation(rec_tpm)

print(f"Feedforward-like system:")
print(f"  System entropy: {h_ff:.3f} bits")
print(f"  Φ proxy:        {phi_ff:.3f}")
print(f"\nRecurrent-like system:")
print(f"  System entropy: {h_rec:.3f} bits")
print(f"  Φ proxy:        {phi_rec:.3f}")
print(f"\nHigher Φ → more integrated → (per IIT) more conscious")

IIT has generated significant controversy. Notably, it predicts that a feedforward network — like a standard deep neural network — has zero or negligible $\Phi$, because information flows in one direction with no integration. A recurrent network has more.

This puts LLMs in a strange position: they're feedforward transformers at inference time.

What Does This Mean for AI?#

Three positions you might hold:

1. Eliminativism about consciousness (Dennett): There are no phenomenal qualia; "what it's like" talk is a confused way of describing functional organization. On this view, AI consciousness is just a matter of sufficiently sophisticated function.

2. Biological naturalism (Searle): Consciousness is a biological phenomenon, requiring the right causal powers of neurons. Software, no matter how sophisticated, no more "understands" than a thermostat.

3. Panpsychism / IIT (Tononi, Chalmers in some moods): Consciousness is fundamental and widespread. Every integrated physical system has some experience; the question is degree and kind.

What I find interesting is that none of these views is obviously right, and AI is making us notice that. When a system like Claude says "I'm not sure whether I have experiences," that answer is philosophically appropriate — not evasive.

A Note on the Phenomenology#

Whatever the metaphysics, there's something worth noticing phenomenologically: the concepts we have for inner life — seeing, understanding, caring, meaning something — were developed in contexts where only biological creatures were in view.

Applying them to AI might be like asking whether prime numbers are heavy. The category might simply not apply — or a new concept might be needed.

"The limits of my language are the limits of my world." — Wittgenstein, Tractatus Logico-Philosophicus

Recommended reading: Chalmers, "Facing Up to the Problem of Consciousness" (1995); Dennett, Consciousness Explained (1991); Tononi, "Consciousness as Integrated Information" (2008).