A comp.ai.philosophy FAQ


Copyright 1999-2001 by Donald R. Tveter, http://www.dontveter.com, commercial use is prohibited. This material cannot be quoted at length or posted elsewhere on the net or included in CD ROM collections. Short quotations are permitted provided proper attribution is given.


How Will We Know When We Have True Artificial Intelligence, Thinking, Creativity, ... ?

A very nice test for determining if an artificial system is equivalent to the human being is to lay out in detail EXACTLY what the artificial system and the human being are doing, then if the two are very close then if the human being is thinking then it is fair to say that the machine is also thinking. When a system is small enough, like a lawn mower or a TV a person can actually look at all the parts and determine how the system works. When the system gets bigger in terms of its number of parts it is not so easy to tell what it is doing.

Currently a common BELIEF is that the human mind operates by using neurons as switches and the mind learns by modifying weights between the neurons. Even if the architecture were to be that simple it would be extremely hard to trace through the workings of the mind on a simple problem like: "How much is 2 plus 2". Since it may never be possible to do even this it is hard to say for sure you know how every aspect of the human mind works. Worse still, neurons are just not that simple, the presence or absence of various molecules in cells ends up altering their performance making it even harder to know EXACTLY how the human mind works. In practice we may never know and this leaves skeptics with an opening to always claim something else is going on so that the computer and the human mind are not essentially the same.

A less common BELIEF at the moment is that the human mind operates quantum mechanically. Here you have even more trouble in determining what the human mind does because you can never be quite sure what the Laws of Physics are. For instance, take gravitation. Newton's Law of Gravity has been known for quite some time but unfortunately it failed to predict the orbit of the planet Mercury. Well, it failed there but nobody threw it out just because of that, they went looking for more sophisticated laws. Einstein came up with General Relativity and explained the orbit of Mercury and as far as anyone can tell RIGHT NOW it gets the right results whenever it has been tested. Unfortunately Einstein's theory seems to be at odds with the other very well tested area of Physics known as Quantum Mechanics. Something seems to be wrong and a new theory of gravity will be needed. Probably someone will find that theory soon and maybe it will work satisfactorily for a hundred or a thousand years and then maybe someone will find a star explodes when it shouldn't and the race will be on to find another theory of gravity. If the human mind is quantum mechanical then knowing how it works depends on knowing what the laws of Physics are, if we can never be sure of them then we can never be sure how the mind works and skeptics will always be able to claim that something else must be going on in the mind that we just don't know about yet. But then there is always the possibility that a quantum mechanical description of how the mind works will turn out to be relatively simple, simple enough that we'll know what it means to "think" and "understand" although that still leaves the loopholes having to do with exactly what the laws of physics are.

There is a slightly weaker test you could use to determine if the artificial system is pretty close to that of the human being: just test its behavior against that of a human being, if the two are pretty close then you'd have a good argument that they're equivalent or effectively equivalent and you've explained human behavior. Here you would show that the robot could do a lot of things people could do: it could walk around, talk intelligently, learn over time, maybe play chess and discuss politics and religion and do many of the typical things people can do. If you could show it behaving very much like a human being then not many people would argue that the robot is not PRETTY CLOSE to being the equivalent of a human being and so you MAY really now understand human thinking and intelligence. Stevan Harnad [Harnad 1989] has called this the Total Turing Test.

The above simple informal test is not very good, more rigorous testing is in order. In Physics, rigorous testing of a theory is fairly easy, the simple formulas predict a value you can attempt to measure. If the value you measure is close enough to the predicted value you have to say the theory is pretty good. But what can you do with robots that at least superficially appear to be the equivalent of human beings? Well you could manufacture large numbers of them, each slightly different in some way, let them roam around the world and see what they do. Will they become teachers? And programmers? And ditch diggers? And artists and musicians? Will some of them turn to crime? Will some become politicians? And if they do engage in all sorts of different occupations how will they compare with their human counterparts? In terms of Art and Music their works might be absolute mediocrities. If they don't produce an occasional genius in some field of endeavor have you REALLY figured out how people think? For them to be the approximate equivalent of human beings then plotting their talent distributions should produce much the same kind of curve that you get from human beings. So here you could actually start measuring numbers including a mean and a standard deviation of the robots' performance. If the numbers come out pretty close to that of the human being then you have a significant argument in favor of the ideas that the machines are equivalent to human beings and that you've actually been able to produce thinking and intelligence in an artificial system.

Now some editorial commentary: no one should go around claiming that they think they understand or have explained human thinking until they have a working system that most people think is PRETTY CLOSE to a human being and which gives the same spectrum of performance as human beings. Unfortunately many people today are excessively confident that they KNOW.

While this test allows at least some numbers to be measured the tests are still being done within a human culture that the robots can imitate. What if they were on their own on a planet all their own? Would you get a Mozart and a Newton and an Einstein? Would you get a Jesus, a Mohammed and a Hitler? Would they develop religions like human beings? But the testing situation gets still worse because this test is the kind of thing you need to run thousands of times so you can get a mean and standard deviation. Plus just because in our human history we have a Mozart, a Jesus, a Hitler, various religions and so on how often do these people and philosophies turn up in human cultures? You'd have to run the human experiment thousands of times to get an idea of what humans are capable of.

Its unlikely that the most extreme of these tests will ever be done but in my opinion if you can test many robots and get talent and performance curves close to that of human beings you are ready to work out the definitions of intelligence, creativity and so on.

A much weaker test was proposed by Alan Turing and so it is known as "The Turing Test". The version of the Turing Test you typically hear about is as follows (the original is more complex). Put a human being or your best computer into a special room and let human judges communicate with the person or computer in the special room via a teletype. After a short period of time have the judges determine whether or not the occupant of the room is a person or a computer. Run this test many times. If the judges decide that a human is a computer 50% of the time and a computer is a human 50% of the time then the computer must also be thinking. Turing's original paper, COMPUTING MACHINERY AND INTELLIGENCE is now online.

The American Association for Artificial Intelligence is preparing web pages on an overview of AI and it includes a section on the Turing Test. The Turing Test Page contains more details and arguments for and against the test. David Chalmers has a collection of online papers on the Turing Test.

A spin-off of the Turing Test is the Chinese Room thought experiment of John Searle. Here you take one human being who does not know Chinese and give them a book full of rules on how to process Chinese text. People submit statements and questions in Chinese and the person in the Chinese room produces statements in response. Let the whole setup be so good that normal people can't tell if the person in the Chinese room is using a book of rules or really is a Chinese person who understands what is submitted to them. But Searle says that the person with the book of rules does not "understand" the inputs so this system of person and rules is not "understanding" Chinese the way that a real person who speaks Chinese "understands". Furthermore the person with the book of rules could be replaced with a collection of water pipes and valves that could also do just as well. This mechanical contraption does not "understand" what it is doing either. Searle then goes on to say that real "understanding" requires something that only real biological creatures can manage thus computers can't "understand" either.

For a longer discussion of the Chinese Room thought experiment together with the usual replies to it see the web page by Larry Hauser: The Chinese Room Argument (Internet Encyclopedia of Philosophy). David Chalmers has a collection of online papers on the Chinese Room Argument.

References

Harnad 1989 Stevan Harnad, "Minds, Machines and Searle", Journal of Theoretical and Experimental Artificial Intelligence, 1989, 1: 5-25.

If you have any questions or comments, write me.

To Don's Home Page