The Turing Imitation Game

The original game

Turing's actual proposal is more specific than the "Turing test" that people usually discuss. The original game has three players: a human interrogator, a human participant, and a machine. The interrogator communicates with both via text, not knowing which is which. The machine's goal is to convince the interrogator it is the human. The human participant's goal is also to convince the interrogator.

Turing proposed that if a machine could play this role successfully, we should say it thinks, not because we've discovered some deep metaphysical fact about machine minds, but because we have no better criterion for thinking in other humans either. We infer minds from behavior. If the behavior is indistinguishable, the inference should be the same.

This was a deliberately operationalist move. Turing was not claiming machines would be conscious. He was arguing that "can machines think?" was a malformed question that should be replaced with a measurable one.

The objections Turing anticipated

Turing listed and answered nine objections in his original paper. Several are still the center of debate.

The theological objection: thinking requires a soul, and machines have no souls. Turing thought this placed an arbitrary limit on what God could endow with a soul.

The consciousness objection: a machine would only be thinking if it were aware of its own thoughts. Turing thought this leads to solipsism: we can't verify consciousness in any other being, human or machine, so it's a poor basis for exclusion.

The heads-in-the-sand objection: people simply don't want to believe machines can think because it's disturbing. Turing noted this was probably the most common reason for resistance, and the least philosophically serious.

The limits of behavioral testing

The imitation game has two prominent critics from within philosophy of mind.

The Chinese Room (Searle, 1980): a system can produce perfectly appropriate outputs without any understanding of what it's processing. Passing the test shows nothing about inner states.

Blockhead (Block, 1981): in principle, a machine could store every possible conversation up to any length and look up responses. It would pass the test without any computation at all, let alone understanding.

Both arguments share a structure: they grant the behavioral criterion and then show it's compatible with the absence of anything we care about. Intelligence, understanding, experience. The Turing test tells you about outputs. It leaves the question of what's producing them entirely open.

What Turing got right is that we don't have access to any other criterion when dealing with other humans either. What his critics show is that this might mean our ordinary attribution of minds is shakier than we thought, not that the machine passes.