AI: Measurement

Modern chatbots imitate intelligent human communication so well that it is often very difficult to understand who you are talking to: a robot or a person. However, they are just imitating, not having a rational conversation. Therefore, to claim that a chatbot has AI is, of course, nonsense. (Kalyaev, I. A. 2024. “How to Measure Artificial Intelligence?”. Sci. Tech. Inf. Process. 51 (5), 422-428)

Yesterday, I asked Claude to review for me the Turing-Church thesis, which proposes an equivalence class of machines and formalisms, any of which are capable of solving any problem that one of the others can solve. For example, the Turing machine, with its infinite one-bit paper tape, can (in time) do anything the Macintosh on which I am typing now can do. Some machines cannot do everything a Turing machine can do: for example, regular expressions can express lots of things, but you can’t use them to count.

I asked Claude about super-Turing machines, machines that could do things that computers (as we know them) cannot. Claude responded with a reasonable objection:

This claim faces a fundamental problem: it is unfalsifiable.

Suppose someone produces a device claiming to be a super-Turing system. How could we verify this claim? We would need to demonstrate that it solves a problem provably non-computable for Turing machines—such as the halting problem or computing Chaitin's Omega. But even if the device provides an answer, how do we verify its correctness? We cannot check it algorithmically, because if we could verify the answer computationally, the problem would be computable after all.

Conversely, we cannot easily refute such a claim either. We might not understand the mechanism; it might be Turing-complete but very fast; and any finite number of test cases cannot prove equivalence.

This matters, because quite a few people want to believe that something in the human mind — consciousness, or ensouledness, or conversability — cannot be computed.

Later, picking up on the title of an old Alan Newell paper, I challenged Claude to a game of twenty questions. It got my desk fan, though it needed all twenty to do it. Of course, it has no experience of desk fans, or any need of one on a hot day, and desk fans are not common topics of conversation.

I also asked it to jog my memory, because in a previous chat it had recommended a specific Iain Banks novel (Excession) that I ought to read, and I’d lost the note.

Finally, I asked Claude to add notes to my book project to record the discussion of Turing-Church and to add a few books (oh dear) to my reading list. Claude did all that, though at first it got the containers wrong because display expressions in the chapter titles confused it. I explained the issue, and we got everything sorted away.

This feels like a rational conversation. I know lots of people who would not be able, at the drop of a hat, to discuss the Turing-Church thesis. Even fewer, having spent 45 minutes or so with Church and Turing, would be up for a discussion of science fiction. Every member of my hometown city council is an actual human, but I doubt any of them would be up to this task. If this is an imitation of rational conversation, it’s a very good one indeed.