Towards A Good Samaritan World

Monday, January 10, 2005


Every now and then you notice that an everyday experience contains a philosophical insight. One struck me the other day: vending machines.

Isn't it frustrating when a vending machine won't accept your dollar? You unfold the corners. You smooth them out carefully. You try another dollar. Sometimes you can have three or four dollars in your wallet and you still can't get access to that candy bar!

A vending machine is an everyday example of artificial intelligence, a phenomenon otherwise easily associated with the mysterious and quasi-magical realms of cutting-edge science or futuristic fiction. Is artificial intelligence possible? Sure: look at the vending machine! It performs an operation-- broadly, selling products; narrowly, perceiving a dollar bill-- which is typically the work of a human intelligence.

But it doesn't do it quite as well as a person would. You wouldn't have to unfold the corner and smooth out the wrinkles and avoid the tiniest tears in order to give a dollar bill to a person.

The vending machine mimics human knowledge of what a dollar bill is. But it doesn't get it quite right. Its definition of a dollar bill is a bit too narrow, too rigid. And this even though a dollar bill is very well-defined as concepts go: dollars are deliberately designed to be highly uniform and easily recognizable in order to better perform their function. With more natural, more plastic concepts, artificial intelligence would face a much more difficult task.

Suppose you were going to design a robot that would recognize an apple. (Apples are more difficult to recognize than dollar bills.)

Whereas dollar bills have a standard, perfect form from which they deviate in varying degrees, apples do not. We could provide the robot with a template of an "ideal apple" from which to measure acceptable deviation, but it would be arbitrary.

The robot would then have to cope with a range of deviation from the norm which would still fall within the concept of "apple." An apple partially eaten is still an apple; but at some point "a slice of apple," or an "apple core," would be more fitting. An apple half eaten should probably still be called an apple, but an apple sliced in two is most certainly not two apples. On the contrary, many apple slices cut from one apple could still be referred to jointly as "an apple." How would the robot deal with that?

Apples come in many sizes and colors and in a more limited variety of shapes, but here a contemporary robot would have a head start: equipped with modern genetic tools, it could extract a microscopic particle of the apple and match up the relevant gene sequences. This is a short cut-- gene sequences are not the way humans recognize apples-- but a functional one. Yet it would have some pitfalls. Take the example of a caramel apple. A robot that snatched the outer layer-- the caramel-- would be deceived into identifying the objects as a non-apple, when a human would quickly put it in the "apple" category. In any case, the robot is perhaps benefiting from a redefinition of terms: in our age of science, we are inclined to defer to rigorous biological taxonomy. We would be willing to acknowledge that some things resembling apples are not members of the apple family.

The most serious problem, though, is that a person would quite easily and naturally point to a pictorial representation of an apple as an "apple." This would baffle the robot. If I point to an apple in a 17th-century Dutch still-life and say, "That's a beautiful apple, it looks delicious," no human being would misunderstand me. The robot, if permitted to take samples, would conclude that the object in question was certainly no apple and bore hardly any resemblance to one.

Plato believed that the real world consisted of shadows, and he postulated a higher world of Forms or Ideas, which real objects merely participated in. While the theory is rather fantastic, it solves a basic problem of communication. A robot could replicate the gene sequence of an apple for the benefit of another robot, but it could not conceive the idea of an apple. For this reason, though a robot could certainly be programmed to, say, distinguish apples from oranges for industrial purposes, it would be forever baffled by humans' habit of speaking of painted apples as apples, and would probably never even be able to negotiate such mundane mysteries as why the uneaten slices of an apple remain one apple together while a partially eaten apple is an apple by itself.

Humans have the property of conceiving ideas. And ideas are essential to communication. Artificial intelligence can mimic certain processes usually characteristic of intelligences, but it cannot conceive ideas, and as a result it will always have the character of a vending machine: you'll have to unfold the corners, smooth out the wrinkles, de-complicate, standardize and sterilize, to be understood by a thing that does not have a mind and does not conceive of ideas.

The deeper question is: what are ideas? Materialists want to reduce them to particles and forces, but have no idea how this could be done. When they tire of blind faith, even fantastic old Plato may deserve a more sympathetic hearing.


Post a Comment

<< Home