Slashdot Mirror
Everything you Want to Know About the Turing Test
An anonymous reader writes "Everything you want to know about
the Turing test provided by
Stanford Encyclopedia of Philosophy. It is their
latest entry."
← Back to Stories (view on slashdot.org)
I think this is the wrong road to go down. We should make computers better for people, but they should be made to actually think and reason. We have enough trouble doing this ourselves and if we have machines doing it they will surely out think us, and then what?
Checkmate
I've always hated the Turing test. It's too subjective, and has forced people into believing that sentience (what the lay-person thinks AI is) can be simulated. It forced AI junkies to think the road to AI was paved by the perfect grammar for English; a pipe dream to be sure.
AI is not being able to have a conversation with your computer, AI is just algorithms -- computing the right answer to complex problems as quickly as possible.
What most people think of as AI is really Artificial Sentience, and the more I learn about computer hardware the more I realize that it will not happen on my PC.
"Times have not become more violent. They have just become more televised."
-Marilyn Manson
That's WAY MORE than I ever wanted to know about the Turing Test!
Assuming you could build such a bot, as soon as you gave it AI you'd be screwed (metaphorically). If the real chick wouldn't sleep with you then any reasonably form of AI wouldn't sleep with you either! :)
--- I used to moderate, then I read the -1 articles and decided having to filter through them was not worth it.
Inventing true computer intelligence (what is often referred to as strong AI), has often been compared to inventing a flying machines by many AI supporters. They claim there were just as many nay-sayers at the end of the 19th century regarding whether we could physically build a flying machine.
I don't remember who, but someone published a great article in Scientific American that claimed the Turing Test has mis-guided the goals of artificial intelligence. He said, instead of trying to build a bird, let's try and build an airplane. Building AI that was truly human-like would be as useless as building a flying machine that was truly bird-like.
The Turing test is not about "cognitive reasoning". Whether or not you "pass" depends on whether or not the "interrogator" (who reads the transcript of a human's conversation with the machine) can tell which participant is the machine. BTW, you find it "ammusing" to know that some humans have failed the Turing test, and some machines have passed it. It really says more about the interrogator and the test than the participants.
I can see it now- while your droid is cooking dinner, DroidSoftCE crashes, and you see the Blue Eyes of Death- flashing, blinking solid bright, electric, full navy blue as he slowly approaches you.... SCARY.
:)
Seriously though: Does WinCE have a BSOD? I've run WinCE quite a bit in the last few years, both as a PDA platform, but more so as a general OS for doing my everyday computing. (Web browsing, programming [on WinCE, not just for it], SSHing, email, IRC, LaTeX) I have had it crash some, but it's actually been quite stable- far fewer crashes for me on my Jornada 720 and iPAQ than I've had with desktop Windows 98 and XP. You'd think XP could manage to keep stable- who knew? For me, WinCE and Win2k have similar stabilities. I'd much rather have WinCE powering my droids than anything else by M$.
Working toward a usable PDA environment in the spirit of Newton OS: Dynapad
Remember back in school when you were asked to define something "in your own words?" The goal was to prevent you from just parroting the definition you got from the book. But most students eventually learn they can change the word order and substitute a few synonyms and still get away with it. The statement you quoted means that doing that doesn't count, since "the dullest of men can do" it: it requires only a basic knowledge of grammar.
...against machine conciousness is that it allows for a perfect simulation of "thinking" by a machine but still says that the machine cannot be judged "concious". In other words, if I am an android that writes poetry, sings and dances, falls in love and runs for the senate and wins etc., I am still not to be considered concious because the stuff between my ears is not of the right type.
At bottom,the question of whether a machine is concious or a "person" is a question of the civil rights we are willing to grant the machine. Searle is often considered a kind of anti-AI bigot for this reason.
Oddly enough (or not) Data's story in Star Trek: Next Generation is probably as good a guide to this issue as you are likely to find.
Cheers,
JHCVH 1
... that humans, when drunk, would easily fail the Turing test? (repetition, shorter sentences, etc...)
No, really.
Searle's Chinese Room argument never makes it past the first room full of undergraduates. Once you outline the scenario, and ask the class, "Ok, does the man following the rules in the Chinese Room know Chinese?" ten hands spring up. The first answer:
"No, but the room knows Chinese."
Duh. I never really understood who takes his argument seriously.
There are no trails. There are no trees out here.
From the link:
In other words, it's a sort of anti-Turing test. I would think that a system using plenty of misspelled words like the above paragraph could easily fool a computer, but is understandable by humans, and could make a good captcha.
<srcasm>Oh great. That's all we need - more spelling mistakes online. Children today are going to grow up not knowing how to spell!</sarcasm>
From a comment below it:
If you were not a fluent english speaker then you would have a great deal of difficulty doing this.
This is also a good point. In fact, every "captcha" I've seen discriminates against some group or other. Anything graphical is impossible for the visually-impaired. But in fact, if you put a lot of spelling mistakes on the page, you will also make it hard for anyone for whom English is a second language. Captchas sound like a neat idea at first, but there's got to be a better way...
I agree, and that's why I want to go to grad school for hard AI. I've seen so many expert systems guys call their products 'AI' that I've lost count. It's not, and I wish they'd stop confusing people. Just because a system 'learns' doesn't mean it's intelligent.
I tend to think the goal of AI research is to produce systems that can learn and react appropriately in different situations that they were never programmed to handle or necessarily anticipate. If that isn't the goal of AI research, what separates it from writing programs on a large scale?
Yeah, you've got it right. Most 'AI' programs out there are your typical Starcraft AIs, the various vision-, speech- and face-recognition software out there, and programs that drive those cool robots around without bumping things. Each program was designed for a specific task, and cannot (by design) grow any larger than that task. This means that none of these programs is really a significant step toward true AI. Some bits and pieces may be salvageable, to be sure.
This post expresses my opinion, not that of my employer. And yes, IAAL.
> When you are building any formal system you have to start with a set of Axioms. If you throw out the Axiom "people think" what do you have to go on? In essence by throwing out the axiom, you are setting up a situation where anything could be considered thinking, because there is no foundation to compare it with.
Science isn't a formal system; it doesn't have axioms. We have to do as best we can simply by looking to see what happens and then trying to understand it.
So we have this notion that "people think", and we have a very vague notion of what "think" means. Where do we go from there? If we start with the notion that thought is something special that can't arise from mechanical processes, we've answered our question by fiat.
But some of us would like to understand how thinking works rather than having an ex cathedra pronouncement that sets it outside of science from the get go. And everything we've learned about the body, the brain, neurons, neurotransmitters, indicates that humans are just big complex machines with no special ingredients. And in the past few centuries we've made marvelous progress at understanding how these components work, with never a need to invoke the supernatural, metaphysical, etc., yet.
So our question for Searle and his ilk is, what the heck is this human "understanding" if not the result of mechanical processes? If it's not the result of mechanical processes, we'd like to see some evidence for that. If it is the result of mechanical processes, why can't it be done in a computer instead of a bag of dirty water?
Searle's argument is just slight of hand to obscure the basic issues, and buying in to his argument requires accepting an 'axiom' that has no empirical support whatsoever, namely that "understanding" is something special that lies outside the mechanical operation of rules. Essentially he assumes his desired conclusion; everything else is just leaves raked over the path to hide its circularity.
Sheesh, evil *and* a jerk. -- Jade
In short, a lot of anti-AI arguments seem to start from the premise that humans are not essentially biological computers
I think this is rather simple to demonstrate (in the strictest meaning of your words, ie. that humans have the inherent limitations of computers as we currently know them) using Goedel's incompleteness theorem: "Within any formal system of sufficient complexity, one can form statements which are neither provable nor disprovable using the axioms of that system."
Computers are perfectly logical, and can acertain truth using only logic. Goedel's theorem tells us that truth is sometimes actually above the scope of logic; that logic cannot demonstrate truth or untruth within a given system of axioms. Humans, on the other hand, can see the truth of a statement even though it is completely unknowable to the logical system in which the statement was formulated [1].
Humans are also capable of easily altering the system of axioms and rules in which we operate. This is completely beyond the capability of modern computers (though perhaps not future incarnations). We may be biological computers for some elevated definition of "computer", but we are certainly above the capabilities of modern computing machines.
[1] one way to discern the truth of unprovable statements in a formal system is to simply add the unprovable true statements to the list of axioms. This adds more incompleteness to the formal system (which can never be eliminated), but (I believe) the unprovable statements become increasingly convoluted. One could thus argue that humans simply have a very long axiomatic list of unprovable truth statements which makes us appear to be above the bounds of traditional logic, but we are still simple logic machines. There would thus be logical statements that would completely dumbfound us forever if they were ever found. I am dubious however.
Higher Logics: where programming meets science.
...was the gom jabbar. The applicant places his hand inside of a pain inducer while a Bene Gesserit witch holds a the gom jabbar to his or her neck. If the applicant removes his or her hand from the box in response to the pain, the highly poisonous and pointy gom jabbar is used and the applicant dies. If the applicant does not remove his or her hand from the box despite the pain, the applicant passes and is considered human. Frank Herbert's theory is that the test of being a human is that a human's intellect allows the human to act in an intelligent manner despite strong animalistic urges to act otherwise. Compared to this, Turing's test seems simplistic - pretend well enough to be a human, and you'll be a human.
/. readers) would not pass as human to the Bene Gesserit. Some may not pass as human to Turing. The question we have to answer before developing a test for intelligence isn't whether or not a computer can be intelligent, but rather what exactly is intelligence. Turing's test is little more than a "if it walks like a duck and talks like a duck, then it's a duck" test. Is that enough? I really don't think that it is. A true intelligence ought to be able to act in an inspired, creative, and perhaps even irrational manner. Many of the things we do are not entirely rational, including much of the partisan discussion concerning various OS's.
It is ironic, however, that a computer would pass the gom jabbar more readily than a homo sapiens. However, both tests start with an implicit principal assumption: A definition as to what a human is. Many of us here (not to single out
144l. ph34r my 133t l3g4l 5k1lz!
We have, in our little calcite skulls, an incredibly advanced technology. So advanced that, for the first 99% of our existence as conscious beings, we simply took it for granted. Then we got thinking about how we think, and the only thing we were equipped to answer with was to say "it's magic." So we posited the idea of a "soul": this nebulous, weightless, insubstantial magic thing that made us who we are, and would live on after the death of our physical bodies.
Slowly, neuroscience has chipped away at the logical need for this magic, even as our desire for its emotional comfort held steady.
I believe our brains are machines. There are perfectly adequate explanations for our thoughts and memories which incorporate absolutely no supernatural mechanisms. Further, positing a supernatural entity which controls our thoughts adds absolutely nothing by way of explanation (any more than simply saying "humans run on magic") while opening up all sorts of uncomfortable logical quandaries: Why would our souls cause us to behave differently when the brain is loaded up with ethanol? Why can people drastically change their personalities after head trauma, strokes, or other brain-related diseases. If a soul can survive physical dissolution of the brain with memories and emotions intact, why isn't it equally unchanging in the face of Zoloft?
Your analysis of the Turing test is quite simply wrong. It's possible--in fact, rather easy--to mimic a passive psychoanalyst as Eliza does. It's even easier to imitate a paranoid schitzophrenic, and easier still to imitate a 12-year old AOL'er. Imitating a normal cocktail conversation would be somewhat more difficult, but still doable. But put a computer up against an intelligent human in a real discussion of ideas, and anything less than true AI is sharkbait.
Part of the problem is, you seem to misunderstand what the Turing test is supposed to be doing. The test, in its most general form, can be used to discriminate between any two sorts of intelligences. A man and a woman imitating a man. A nuclear scientist and someone pretending to be a nuclear scientist. A paranoid schitzophrenic and a computer pretending to be a paranoid schitzophrenic.
If I were to build a machine that imitated your friend Buddy, the Turing test would be to put you in front of two screens, one with the real Buddy and the other hooked up to my machine. If you were only able to guess which was Buddy half the time, my machine would not only have passed the broader Turing test (which only says that the respondent is intelligent), but you would also have to admit that the machine was substantially similar to Buddy's mind.
Your snippet of conversation is proof of your misunderstanding. Any computer can fool a sufficiently oblivious person into thinking they're having a conversation. Where the tread hits the tarmac is when an intelligent person, looking for signs of non-intelligence and fails to find it. A real Turing conversation would go something like:
You want the truthiness? You can't handle the truthiness!
The problem with this type of test is (as stated in one of the comments below) you are only able to actually pass it if you are a native speaker. This becomes very obvious if you have ever tried tutoring non-English speakers in English. A Japanese student, who's spelling was amazingly good, much better than mine, was completely unable to read a sentence if only a single word in that sentence was misspelt. It also worked the other way around. I have a (very) basic grasp of Japanese (thanks to said student) but if 1 letter was wrong there was no way I would be able to understand any of what was written.
There's also the problem that most people assume thinking is synonymous with the subjective perceptual experience. That the experience of blue you have when looking at the sky, say, is inextricably intertwined with thought.
I can easily envision a machine capable of thought exactly as we do without that machine necessarily being subjectively "alive".
Until physics comes up with a way to explain the very real subjective experience that I, and probably most of you, have, there'll be a big chunk missing.
I can envision a machine emulating atoms and whatnot, including an entire brain, digitally, to any desired physical degree. Flip the switch, and the "brain" would interact just as a human's would. The $64,000 question is would it have a subjective perceptual experience? Would such a thing be required for thought? Would it arise out of the pure information interaction, as some suggest, or is it something peculiar to our particular molecular machine, as Searle seems to think?
"Has [being a kidnapped teenage girl, raped repeatedly for months] changed you?" - Katie Couric to Elizabeth Smart
Forgive me. I posted this in a different article a few days ago. I didn't really get any feedback. This subject fits my question much better.
Lets say that one day a computer passes the Turing test. Then lets say that a few years later it passes it again, only this time it passes utilizing voice recognition and speech synthesis.
If you think about it this is gonna be a really hard test for the computer to pass. I can't even imagine what is involved in figuring in voice inflection, accents and stuff like that.
Anyways it is at the point where you are on the telephone and you can converse with a computer and you have no idea it is a machine.
Hence, it passed the Turing test.
What happens if the computer begins to make the argument that turning it off and disasembling it is no different then killing a person or an animal?
What happens if the computer starts to make the argument that it is capable of thought?
What are the implications of that?