Must AIs be embodied?
Cognitive Daily has an interesting discussion going on about whether or not an artificial intelligence requires a body. There are some interesting posts in the discussion, but as with most of these discussions, it quickly turned into a matter of “what does it mean to be intelligent?”
To drop my two cents in on the matter, I must say that the answer to the question is obviously yes; but we have to define “body.”
If you define body to be a physical object in the world that is capable of moving around and interacting with reality, the answer is no. If, on the other hand, you define “body” simply as “a means of obtaining input for the mental processes to act on,” then by all means the answer is yes.
My point is that, while intelligence needs some sort of sensory input to act on (else what the hell would it mean for it to be intelligent?), we shouldn’t be physical chauvinists about where those sensory inputs come from. They can be byte streams from a network interface, the pixels of a camera, or the electrical signals from a piece of meat sensitive to touch.
Now, we can modify this question and ask “can an artificial intelligence have human like intelligence without a body?”
One comment in particular seems to answer this formulation of the question in the negative:
“Can an AI understand the written or spoken word “ball” without associating it with the visual patterns of seeing various balls, the feeling (real or virtual) of holding one, the intuitive physics of throwing, bouncing, and catching one, the concept of games, other spherical objects, etc.?
….
I believe that computational AI is possible (along with machine consciousness) and that many tasks expected of an AI may be possible with limited sensory connectivity, but some tasks necessary, say, to pass a Turing Test might require a “body” (real or virtual, humanoid or abstract) with a full (or augmented) set of senses in order to understand the depth and nuance of language, not to mention human behavior, art, humor, etc.”
Well, why couldn’t it understand “ball”? Don’t we all claim to understand words associated with objects that don’t, or in some cases couldn’t really exist? We sometimes talk about square circles, even though the concept is a logical (and physical) contradiction (boxing rings to the contrary), but that doesn’t seem to stop us from thinking we understand what the (non-existent) referent of the word is.
And what about something like Grahams Number, which I mentioned in my previous post? Does anyone really know what Graham’s number is? Certainly not in the same sense that we “know” what 3 is. But this doesn’t stop us from intelligently talking about it, reasoning about it, ect.
You could argue that we are only able to do so because we have interacted with other things in a physical way, and are extrapolating from the existent objects to the non-existent ones. I have some understanding of Grahams Number, because I already posses understanding of numbers in general.
What it boils down to is a claim that to get to a “understand the depth and nuance of language, art, etc.” requires you to get to the understanding in a certain way, i.e. by interacting with a sensory environment of a certain type (virtual or actual). This is an empirical question that we cannot solve from the comfort of our armchairs, but my gut feeling is that the claim is wrong. Violating my previous statement, from the comfort of my armchair, I fail to see why we can’t start with, say, just raw logic or some other incredibly weird sense data and arrive at the same place that humans do via another path.
But, then, what are we to think of people who are truly color blind, as in they see only monochromatic colors? Do we want to say that they could not possible understand the word “color”? Or how about someone who was born a blind paraplegic? Could they ever truly appreciate the word “ball” since they have never seen nor felt a spherical object, or watched the physical actions of objects? Can a blind person really understand what a triangle is? Interesting questions to which I do not have the answer, though I’m inclined to say yes, with the qualification that their understanding obviously won’t be exactly the same as ours, but that doesn’t make ours “right” and theirs “wrong.”
As to the second part of his comment, I’m not entirely sure I buy it. Certainly Good Old Fashioned AI (GOFAI), the idea that we could just program a human grade artificial intelligence, turned out to be much harder then initial estimates. But just because it is hard does not mean it is impossible. We’ve never logically proven that you cannot write an imperative computer program that would pass the turing test, so the jury is still out on whether or not it is possible. A mountain of empirical evidence in the form of “every attempt so far has failed” doesn’t mean it is impossible.
How could we show that it is impossible? By showing that the brain computes a non-computable function would do nicely. For attempts in this direction, consult J. R. Lucas Minds, Machines, and Godel and Mechanism: A Rejoinder.
I’ll post a review of Lucas’s position in a future post.
One last, small tangent. Someone posted a lengthy reply on their own blog, where they also make many interesting points, but fall victim to a common misunderstanding regarding the Turing Test. Specifically, many people, both inside and outside the AI community, seems to think that the turing test is an intelligence test. I.e. if you cannot pass the turing test, then you are not intelligent.
This is absolutely wrong, as I think is clear if you read Turing’s original paper. The test is only meaningful in the positive, and not the negative, aspect. If you can pass the turing test, we must grant that you are intelligent. On the other hand, if you fail the turing test, we must do nothing. Failing the turing test is not grounds for denying intelligence, but passing it IS grounds for asserting intelligence.