Home / Philosophy of Mind

The Hard Problem of Consciousness and AI

A glowing mechanism of channels with an empty dark centre

The hard problem of consciousness is the question of why any physical process is accompanied by experience at all. For AI it means this: even if we could explain everything an AI system does, from how it sorts information to what it says about itself, we might still not know whether there is anything it is like to be that system. That gap is the deepest reason the debate about AI consciousness is so hard to close.

What the hard problem of consciousness is

The philosopher David Chalmers gave the problem its name in 1995, in "Facing Up to the Problem of Consciousness", published in the Journal of Consciousness Studies. He wrote that the really hard problem of consciousness is the problem of experience. The Internet Encyclopedia of Philosophy puts it as the problem of explaining why any physical state is conscious rather than nonconscious.

Chalmers set it against what he called the easy problems. These are not easy in practice; they are "easy" because they seem open to the ordinary methods of cognitive science. His examples include:

  • telling stimuli apart, sorting them and reacting to them;
  • integrating information;
  • reporting mental states;
  • a system accessing its own internal states;
  • focusing attention and deliberately controlling behaviour;
  • the difference between being awake and asleep.

Each of these is a function, and functions can be explained by describing a mechanism that performs them. The hard problem is that, once every function is explained, it still seems to make sense to ask why performing them feels like anything.

An open music box with a glowing tendril of light curling up

The hard problem of consciousness and AI systems

AI makes the problem concrete. Several of Chalmers's easy problems describe things AI systems are now built to do. A language model integrates information, reports on its states when asked, and in some experiments can detect changes to its own internal activity, as described in how to test if AI is conscious. Solving easy problems in a machine, on this view, does not tell us whether the hard one has been solved there too.

That is why theories matter so much. 80,000 Hours notes that several theories, including functionalism, global workspace theory, higher-order thought theory and integrated information theory, could allow digital minds to be conscious, while biological theories and some forms of dualism could rule them out. Which side a machine falls on depends on which theory you accept, and the hard problem is the reason none of them has been proved.

A closer look: the main responses

The Internet Encyclopedia of Philosophy groups the responses to the hard problem. Applied to AI, they give very different answers:

  • Eliminativism: there is no further fact of experience to explain, so the question about AI reduces to what it does.
  • Strong reductionism: experience is fully explained by function, as in global workspace and higher-order theories; an AI with the right functions would be conscious. See global workspace theory and AI.
  • Weak reductionism: experience is identical to certain physical states, though we cannot say why; whether a machine qualifies depends on its physical make-up.
  • Mysterianism: the problem is unsolved, perhaps permanently, so we may never know about AI.
  • Dualism and epiphenomenalism: experience is non-physical, which leaves open whether a machine could have it at all.
  • Dual-aspect theory, neutral monism and panpsychism: something more basic underlies both matter and experience; panpsychism holds that everything has some degree of experience.

The Chinese room argument makes a related point about meaning: running the right program may not be enough for understanding.

Checklist: spotting the hard problem in an AI claim

  • Is the claim about a function or about experience? "It reports its internal state" is an easy-problem claim.
  • What theory bridges the gap? Any move from function to experience assumes one of the responses above.
  • Is uncertainty being admitted? Even researchers who doubt current systems, like Chalmers, who judged current language models probably not conscious, leave the future open.

More in Philosophy of Mind · Report a correction