radzy@radzy.UUCP (Tim Radzykewycz) (02/03/90)
In all the discussion I've seen about the Turing test, people have often brought up the possibility of asking the question "Are you human?" In all the other discussions (e.g. other than this one on the net) SOMEBODY has always brought up the point that the HUMAN in the test might also lie. I guess I have to be the one this time. How can the person asking the question know for certain that the answers of either subject (e.g. human or computer program) are true? What are the criteria which s/he can use to determine this? If you were the one asking the questions, would you ask this question and base your decision on the result? What about the possibility of a program which simply printed one of "yes", "no", and "maybe" at random after receiving any input? This discussion has gone too long looking only at the computer program side of things. You also have to look at the other side, the one with a human behind it. That's my humble opinion, anyway. -- Tim "radzy" Radzykewycz The Incredible Radical Cabbage radzy@cogsci.berkeley.edu - or - radzy@radzy.net.com
janeric@control.lth.se (Jan Eric Larsson) (02/05/90)
In article <376@radzy.UUCP> radzy@radzy.PacBell.COM (Tim Radzykewycz) writes: >How can the person asking the question know for certain that >the answers of either subject (e.g. human or computer program) >are true? What are the criteria which s/he can use to determine >this? In the Turing test (as described by Turing) the computer/program is supposed to "pose" as a human, by giving all sorts of answers. Obviously, most of these will be lies. It is important to observe that the Turing test is no test for intelligence or consciousness, at least not according to Turing. It would only prove that the computer is good at "posing" as a human. The good Alan would certainly not agree with either Searle or the defenders of the infamous "strong AI", when it comes to the conclusions of the Turing test. See Turing's article in Mind, october, 1950. Jan Eric Larsson JanEric@Control.LTH.Se +46 46 108795 Department of Automatic Control Lund Institute of Technology "We watched the thermocouples dance to the Box 118, S-221 00 LUND, Sweden spirited tunes of a high frequency band."
ruth@aiai.ed.ac.uk (Ruth Aylett) (02/05/90)
As I recall, deception was an essential element of the original Turing Test - the one that involved a man and a woman. The task was to distinguish which was the real woman, given that the man was doing his best to appear as a woman too. To stand a chance of succeeding, the man has to possess the human capacities of imagination, empathy, and understanding of tests and games, that is social ritual. Would any intelligent entity (assuming we can imagine what a non-human intelligence might be like) be wedded to the literal truth? Or, in the original case, why would the man want to do something as ridiculous as pretend to be a woman over a teletype? As a matter of interest, has anyone anywhere ever tried this original turing test? If so, with what result? Ruth Aylett ruth@aiai.uucp R.Aylett@uk.ac.ed
kp@uts.amdahl.com (Ken Presting) (02/06/90)
In article <376@radzy.UUCP> radzy@radzy.PacBell.COM (Tim Radzykewycz) writes: >In all the discussion I've seen about the Turing test, people >have often brought up the possibility of asking the question >"Are you human?" In all the other discussions (e.g. other than >this one on the net) SOMEBODY has always brought up the point >that the HUMAN in the test might also lie. I guess I have to >be the one this time. > >How can the person asking the question know for certain that >the answers of either subject (e.g. human or computer program) >are true? What are the criteria which s/he can use to determine >this? This particular problem is easy. If you ask a few questions like "What city are you in?" and "What's the weather like?" you can use a newspaper to tell whether the answers are consistent. The idea is to use yet another instance of a task which is trivial for people but tedious for computers. Almost every discussion of the Turing Test I've read focuses on the subjective experience of humans, and whether a computer can give a convincing impression of having the same experiences - hopes, fears, etc. This is perhaps a natural place to focus, if we suppose that the computer will have only a teletype with which to interact with the outside world. It might seem excessively "unfair to the machine" to expect it to be prepared with up-to-the-minute data on weather or current events. Look at how easy it would be to trip up a computer in a discussion about daily life. People ate dinner last night, went out, or watched TV, or read a newspaper, and in the process acquired some topical information. For a computer to convince us that it occupies its evenings similarly, a data entry operation of massive proportion would be required. This is my favorite reason for supposing that the Turing Test will NEVER be passed. No funding agency in its right mind would underwrite the project. Of course, it's no slur on an AI that it doesn't watch Johnny Carson, or talk about the weather. Ian Sutherland comments: > . . . Let's try and come a little >closer to passing the Turing test before we start worrying about >whether passing the Turing test is adequate or not. We might be able to specify an objective criterion which is easier to implement, yet more convincing than the TT. One aspect of the TT that reduces its significance is that a machine can pass it without knowing it's a machine. Let me emphasize this: LYING IS NOT A PROBLEM. NOT KNOWING THE TRUTH IS. An elementary ability to make descriptive statements about one's own body and its current condition is present in children soon after they learn to speak. Animals can't talk, but they clearly adapt their behavior to their various states of hunger, temperature, etc. Absence of this ability completely precludes any claim that the system in question is conscious.
weyand@csli.Stanford.EDU (Chris Weyand) (02/06/90)
References: <15439@well.UUCP> <11673@csli.Stanford.EDU> <11324@venera.isi.edu> <1700@castle.ed.ac.uk> <11489@venera.UUCP> <6340@sdcc6.ucsd.edu> <7cHZ028I81fo01@amdahl.uts.amdahl.com> <4819@convex.convex.com> <376@radzy.UUCP> <f3Gj02be85fI01@amdahl.