Blockhead

The scenario

Suppose you precomputed every possible conversation up to some length, say a million turns. You stored every possible input a human interrogator might type, paired with an appropriate human-seeming response. The machine receiving input simply looks up the response and returns it. No computation. No reasoning. No understanding of any kind. Just a lookup.

This system passes the Turing test. Given any input, it returns a convincing response. The interrogator finds nothing unusual.

Block calls this system "Blockhead." It is the most minimal possible thing that could satisfy the Turing test, a pure input-output table with no interior whatsoever.

Why the Turing test is insufficient

The Turing test was designed to sidestep unanswerable questions about inner experience. If the behavior is right, we say the machine thinks. Blockhead shows this move fails.

Blockhead's behavior is not just similar to intelligent behavior. It is identical to it, at the level of outputs. If the Turing test were a sufficient criterion for intelligence, Blockhead would be intelligent. But nothing is happening inside Blockhead that resembles thought. It is a filing cabinet with very good indexing.

The test only measures what comes out. It says nothing about the process that produced the output, and the process is where thought actually lives, if it lives anywhere.

What this reveals about intelligence

Blockhead separates two things that are easy to conflate: producing intelligent-seeming outputs and being intelligent.

We normally treat these as correlated because, in humans and animals, there is no gap between them. A human who gives consistently apt, creative, contextually sensitive responses is, by any reasonable standard, thinking. But that correlation holds because of how human intelligence works, not because outputs and intelligence are the same thing.

Blockhead breaks the correlation by stipulation. The outputs are there; the intelligence is not.

This matters beyond the Turing test. Any behavioral criterion for mind faces the same problem. If you specify a behavioral test precisely enough, you can imagine a system that satisfies it purely through pre-compiled responses, without anything that functions as understanding. The outputs are the shadow of thought, not thought itself.