- What is understanding? What is consciousness? What is meaning? What does it mean to think? These, of course, are philosopher's questions. They are the bread and butter of philosophy. But what of the role of such questions in AI? Shouldn't AI researchers be equally concerned with such questions? I believe the answer to be yes and no.
- According to the distinction between weak and strong AI, I would have to place myself in the weak AI camp with a will to move to the strong side. In a footnote, Searle mentions that he is not saying that I am necessarily committed to the two "AI claims" he cites. He states that claims that computers can understand stories or that programs can explain understanding in humans are unsupported by my work.
- He is certainly right in that statement. No program we have written can be said to truly understand yet. Because of that, no program we have written "explains the human ability to understand."
- I agree with Searle on this for two reasons. First, we are by no means finished with building understanding machines. Our programs are at this stage partial and incomplete. They cannot be said to be truly understanding. Because of this they cannot be anything more than partial explanations of human abilities.
- Of course, I realize that Searle is making a larger claim than this. He means that our programs never will be able to understand or explain human abilities. On the latter claim he is clearly quite wrong. Our programs have provided successful embodiments of theories that were later tested on human subjects. All experimental work in psychology to date has shown, for example, that our notion of a script (Schank & Abelson 1977) is very much an explanation of human abilities (see Nelson & Gruendel 1978; Gruendel 1980; Smith, Adams, & Schorr 1978; Bower, Black, & Turner 1979; Graesser et al. 1979; Anderson 1980).
- All of the above papers are reports of experiments on human subjects that support the notion of a script. Of course, Searle can hedge here and say that it was our theories rather than our programs that explained human abilities in that instance. In that case, I can only attempt to explain carefully my view of what AI is all about. We cannot have theories apart from our computer implementations of those theories. The range of the phenomena to be explained is too broad and detailed to be covered by a theory written in English. We can only know if our theories of understanding are plausible if they can work by being tested on a machine.
- Searle is left with objecting to psychological experiments themselves as adequate tests of theories of human abilities. Does he regard psychology as irrelevant? The evidence suggests that he does, although he is not so explicit on this point. This brings me back to his first argument. "Can a machine understand?" Or, to it put another way, can a process model of understanding tell us something about understanding? This question applies whether the target of attack is AI or psychology.
- To answer this question I will attempt to draw an analogy. Try to explain what "life" is. We can give various biological explanations of life. But, in the end, I ask, what is the essence of life? What is it that distinguishes a dead body that is physically intact from a live body? Yes, of course, the processes are ongoing in the live one and not going (or "dead") in the dead one. But how to start them up again? The jolt of electricity from Dr. Frankenstein? What is the "starter"? What makes life?
- Biologists can give various process explanations of life, but in the end that elusive "starter of life" remains unclear. And so it is with understanding and consciousness.
- We attribute understanding, consciousness, and life to others on the grounds that we ourselves have these commodities. We really don't know if anyone else "understands," "thinks," or even is "alive." We assume it on the rather unscientific basis that since we are all these things, others must be also.
- We cannot give scientific explanations for any of these phenomena. Surely the answers, formulated in chemical terms, should not satisfy Searle. I find it hard to believe that what philosophers have been after for centuries were chemical explanations for the phenomena that pervade our lives.
- Yet, that is the position that Searle forces himself into. Because, apart from chemical explanation, what is left? We need explanations in human terms, in terms of the entities that we meet and deal with in our daily lives, that will satisfy our need to know about these things.
- Now I will return to my analogy. Can we get at the process explanation of "life"? Yes, of course, we could build a model that functioned "as if it were alive," a robot. Would it be alive?
- The same argument can be made with respect to consciousness and understanding. We could build programs that functioned as if they understood or had free conscious thought. Would they be conscious? Would they really understand?
- I view these questions somewhat differently from most of my AI colleagues. I do not attribute beliefs to thermostats, car engines, or computers. My answers to the above questions are tentative no's. A robot is not alive. Our story-understanding systems do not understand in the sense of the term that means true empathy of feeling and expression.
- Can we ever hope to get our programs to "understand" at that level? Can we ever create "life"? Those are, after all, empirical questions.
- In the end, my objection to Searle's remarks can be formulated this way. Does the brain understand? Certainly we humans understand, but does that lump of matter we call our brain understand? All that is going on there is so many chemical reactions and electrical impulses, just so many Chinese symbols.
- Understanding means finding the system behind the Chinese symbols, whether written for brains or for computers. The person who wrote the rules for Searle to use to put out the correct Chinese symbols at the appropriate time - now that was a linguist worth hiring. What rules did he write? The linguist who wrote the rules "understood" in the deep sense how the Chinese language works. And, the rules he wrote embodied that understanding.
- Searle wants to call into question the enterprise of AI, but in the end, even he must appreciate that the rules for manipulating Chinese symbols would be a great achievement. To write them would require a great understanding of the nature of language. Such rules would satisfy many of the questions of philosophy, linguistics, psychology, and AI.
- Does Searle, who is using those rules, understand? No. Does the hardware configuration of the computer understand? No. Does the hardware configuration of the brain understand? No.
- Who understands then? Why, the person who wrote the rules of course. And who is he? He is what is called an AI researcher.
Text Colour Conventions (see disclaimer)
- Blue: Text by me; © Theo Todman, 2019
- Mauve: Text by correspondent(s) or other author(s); © the author(s)