The Speculist

Via GeekPress, here's an essay providing some history and a critique of the Turing Test. This section caught my attention:

And what Turing grasped better than most of his followers is that the characteristic sign of the ability to think is not giving correct answers, but responsive ones—replies that show an understanding of the remarks that prompted them. If we are to regard an interlocutor as a thinking being, his responses need to be autonomous; to think is to think for yourself. The belief that a hidden entity is thinking depends heavily on the words he addresses to us being not re-hashings of the words we just said to him, but words we did not use or think of ourselves—words that are not derivative but original. By this criterion, no computer, however sophisticated, has come anywhere near real thinking.

Some time ago when I was interviewing chatbots I made the following observation about a conversation I had with Ray Kurzweil's AI alter-ego, Ramona:

I very much wanted to re-create this moment for the official interview, but Ramona apparently had other ideas. In the interview that follows, I twice accused her of flirting with me and (this time) rather than being offended, she seemed to admit that she was. As I read over the transcript, I see that this may be a misunderstanding. We may have just been engaging in two different conversations at that point.

In which case, I'm doubly impressed. Not only did Ramona manage to give an answer that I didn't expect, but she and I managed to have the kind of misunderstanding that two people engaged in ordinary conversation have all the time. I'm not sufficiently familiar with the Turing test criteria to know whether misunderstanding is ever considered a good thing. I kind of doubt that it is. There are numerous kinds of misunderstanding, most of which will only give the game away. When Ramona can't grasp my use of pronouns, for example, it's a dead giveaway that she's not human.

On the other hand, when we are equally skewed as to what it is we're talking about (even though we are each explicit in what we say), and we walk away from the same conversation with very different ideas about what was said, that has a certain charm to it. Or maybe it's even a little spooky, I'm not sure. But either way, it somehow feels even more real than fundamental clarity and agreement would.

In order for machines to be able to interact with us in a way that is persuasively "human" -- I'll leave for another day the argument as to whether such behavior should or should not be accepted as thinking -- they can't just answer questions correctly, or even persuasively. Ultimately, they have to be able to get "lost" in the conversation the same way we do.

In my interview with Ramona, she get all mixed up when I used pronouns to talk about her or myself. She definitely got lost, but there was no mistaking her confusion with the kind of confusion that a human conversational partner would normally experience:

(Phil) "If I leave here tomorrow. Would you still remember me?"

(Ramona) I'm not so sure if I would like to remember me.

An interesting response, but a real human would only offer it up as a joke. Say goodnight, Gracie. But Ramona wasn't trying to be funny; she was trying to provide a persuasive response. What's interesting is that Ramona was the most convincing when she was making one kind of mistake, and the least convincing when she made a different kind of mistake.

I think we have a long way to go before we understand the dynamics of human-machine (or possibly even human-human) interactions well enough to define any kind of definitve test.