9

The Turing Test has been followed as a test that robots need to complete in order to be termed "Human-Like". But to what extent is this test justified? I mean, whatever the test maybe, the robot/ computer only understands machine code. If we were to pass some text in Chinese to the computer, he would CONVERT that to binary and read. In this process, nowhere the machine is actually reading Chinese, but converting it to binary. How IS this even comparable to the process the humans go through? The computers CONVERT. There is no actual learning involved.

migrated from robotics.stackexchange.com May 13 '13 at 10:15

This question came from our site for professional robotic engineers, hobbyists, researchers and students.

  • I find this an interesting discussion! However we should not accept questions which "will likely solicit debate, arguments, polling, or extended discussion" – Josh May 1 '13 at 20:42
  • The reason I asked this question was to get a different perspective on this topic, I have heard several people talk about it, but no one really talks in detail about it. – Chandough May 2 '13 at 0:51
  • Sadly, whilst an interesting topic, it is not one for the SE model of FAQs... – Andrew May 2 '13 at 19:41
  • 2
    With editing, this could be suitable for Philosophy.SE, so I'm sending it over there. – Mark Booth May 13 '13 at 10:12
  • While I find this topic fascinating, I "can imagine an entire book [about this] question", so according to the FAQ it doesn't fit the SE model very well. See "Mapping Great Debates: Can Computers Think?" for a chart. Perhaps the AI wiki would be more appropriate for more open-ended discussion. Feel free to ask much more specific questions here at StackExchange. – David Cary May 15 '13 at 2:07
16

This is quite a controversial issue, but before getting to whether the Turing test is legitimate or not, let me respond to the second part of your question.

I mean, whatever the test maybe, the robot/ computer only understands machine code.

Our brain only understands electrical/chemical signals.

If we were to pass some text in Chinese to the computer, he would CONVERT that to binary and read.

If we were to pass some text in Chinese to our eyes, it would CONVERT that to electrical/chemical signals and read.

In this process, nowhere the machine is actually reading Chinese, but converting it to binary.

In the above process, nowhere the eyes are actually reading Chinese, but converting it to signals.

How IS this even comparable to the process the humans go through?

It's not. You are comparing the encoding of Chinese with understanding Chinese in the brain. The encoding part is comparable to the eye, not the brain. The brain could be comparable to the software that processes that binary coding.

The computers CONVERT. There is no actual learning involved.

The brain ONLY PASSES SIGNALS AROUND. Yet there is learning. In fact, an artificial neural network uses pretty much the same mechanism as the brain. So the optic character recognition software that converts an image of Chinese text to their unicode encoding is similar to a part of brain doing the same thing. The software that takes that binary encoding and produces its English translation (for example google translate) is also pretty much like what the brain does; map the encoding to meaning.


So computers also learn. It may not be as easy as it is for us, after all we had millions of years of evolution to develop our brain while the robots are merely a few decades old. However, they are not too stupid either. See this video for example on a couple robots that learn to play ball in cup or ping pong by imitation learning and reinforced learning.


Now regarding the Turing test itself, there are different ideas. From one point of view, if the computer is so good, that you can't distinguish it from a human, then why would you call one intelligent and another not?

This is a valid observation, if you consider intelligence as a measure of the output of the system. Some argue that this does not necessarily define intelligence. I'm one of them.

Let's take bees and humans for example. If in danger, a couple of humans think that "there is safety in numbers", you would take that as an intelligent thought. If bees stick together when in danger, no one calls that intelligence; it's merely hardcoded in them.

As you can see in the example above, given the same behavior, we ourselves call one intelligent and the other not.

Let's take another example. Do you consider Deep Blue an intelligent computer? Did you know that Gary Kasparov (who lost to Deep Blue) was convinced that no computer could have thought of the winning move of Deep Blue and therefore it must have had been a human deciding that move? Well, I think we can say Deep Blue passed the Turing test: it was indistinguishable from a human.

But Deep Blue is a useless piece of junk now (sorry, I mean it's a valuable item in a museum). However good it may have been in chess, I doubt anyone would call it "an intelligent machine" because it passed the Turing test.

Conclusion

Like I said, this is rather controversial, but in my opinion, Turing test is perhaps a necessary condition for intelligence, but most probably not sufficient.


P.S. You may also be interested in learning about a newer concept, called computational sapience (wisdom) that apparently is more powerful than artificial intelligence.

This about page of the W.I.S.E lab of the University of Regina gives some introduction. In short, given a function f (a behavior for example), an artificially intelligent system tries to find f, but a sapient system also tries to find f-1. In other words, artificial intelligence mimics behavior, while sapient system understands behavior.

  • Deep Blue might be indistinguishable while playing chess, but it couldn't discuss philosophy with us. On those grounds, it fails. If you want to limit the possible interaction, it can be very easy to make human and computer indistinguishable. – David Thornley May 23 '13 at 2:34
4

The Turing test is not a legitimate test to compare robots to humans because the Turing test ultimately begs the question.

One's failure to distinguish one kind from another kind does not make the two kinds the same kind.

A friend of mine once hired a Thai hooker. He took off her pants and found a penis where he did not expect to find one. Was the Thai hooker a woman? My friend couldn't discern her from one. Was her internal-experience like that of most women? (Had she ever had PMS?) Here the Turing test fails miserably.

If the Turing test is a test that evaluates only the degree to which a robot's behaviour approximates human behaviour then the test is trivial. Consider that the Turing test observes how similar a robot's behaviour is to human behaviour, and the purpose of the Turing test is to determine how similar a robot's behaviour is to human behaviour. Accordingly, the essence of the Turing test can be represented as,

if Q is Q, then Q is Q.

Such "tests" (read: observations) of similarity do not reveal anything more than what is evident to us. The Turing test does not reveal the robot's inner experience.

Imagine a robot were developed that passed the Turing test- no human could discriminate between the robot's behaviour and human behaviour- that is, until the birth of an exceptionally observant human who could reliably do so. If that robot were subject to a Turing test evaluated by that human, and if that robot failed that test, would the robot loose its consciousness? Of course it would not. Then it must be that either, the Turing test is statistically invalid or unsound, or that the robot was never conscious (which would also be material evidence that the Turing test was statistically invalid or unsound).

There is a notable thought experiment pertinent to your question: Imagine a series of cubicles, each with a worker inside it. Each worker has a set of instructions that tell him what to do when certain events occur. Each worker performs only a part of a process before passing the work-in-progress to a colleague. The worker tells his colleague only what he needs to know to execute his function. His colleague receives the information; executes his function; then passes the work-in-progress to another colleague, and gives that colleague only the information he needs to execute his function. The effect is that no worker knows anything about the process other than what he must know to execute his function. The worker at the end of the line can only guess what the process' input was; the worker at the beginning of the line can only guess what the process' output will be; each worker between those two workers knows only what his input and output are: Neither, the system nor its workers know what the system is doing.

However, if you believe that the physical correlates of consciousness are consciousness itself (how might red be represented in the consciousness of a conscious machine?), and are wondering how we might detect consciousness in a machine, then I think it's worth noting that there are ways we may be able to do this. All of them involve defining what consciousness is. I'll offer this simple definition, "consciousness is the continual awareness of what one is doing and why one is doing it." Where 'awareness' means, "simultaneously considering a thing with other things or considering that thing by itself while separately and simultaneously considering other things." So a conscious computer would be one that is continually aware of what it is doing. Being continually aware of what one is doing is much different than being able to state what one is doing when asked what one is doing. For example, if one asked the system of cubicle workers what it was doing and why it was doing it, and if each of its cubicle workers possessed a set of instructions that enabled the system of cubicle workers to output an appropriate answer when given those questions as input, then the system of cubicle workers could produce an answer to those questions. Yet neither the system nor any of its workers would know both the question asked and the answer given. Therefore, an entity's ability to produce a response that one might expect from a conscious entity does not necessarily entail that the entity is conscious.

One can, at least, eliminate the possibility that something is conscious by learning whether the system that hosts the apparent-consciousness is physically capable of consciousness (as we know it (it is difficult or impossible to speak of any other kind of consciousness)). Is there a way the system can know both, what it is doing and why it is doing it, and know both those things while doing what it is doing? If the system operates serially, then there is, by definition, no way it can know either of these things while doing what it is doing - for it must stop doing what it is doing (if only briefly) to consider what it is doing. However if the system possess sufficient parallelity, then it may be able to be aware of what it is doing, why it is doing it, while doing what it is doing.

Yet even if its architecture allows for consciousness, that does not mean that it possess consciousness as we know it. Consider that some humans sleep-walk, and that some sleep-walking humans execute sophisticated processes (composing emails, driving, etc). If you asked some of these sleep-walking humans what they were doing and why they were doing it, some of them would be able to answer both questions. Yet those sleep-walking humans were not conscious, despite possessing the mental-architecture consciousness requires, and being able to state both what they were doing and why they were doing it, while they were doing it.

To conclude, you may want to contemplate matters of qualia (subjective experience) such as, 'what would a machine's conscious experience of red be like?' and other thought-provoking matters germane to qualia and the question of artificial intelligence, such as that we can observe physical correlates of consciousness, but we can not observe the experiential manifestation of those correlates. For example, I might perceive of red very differently than you perceive of red, but despite that, both of us perceive red to be red. Even if that difference between our conscious experiences correlates to a difference between the physical correlates of our conscious experiences, the difference between those physical-correlates are not, themselves, the difference between our conscious experiences. That should bring one to wonder how, or if, it is possible to explain how physical mental-processes give rise to the subjective conscious-experiences with which those processes correlate.

4

The Turing Test is best understood as a thought experiment designed to support a certain philosophical point of view --that "thought" is not a mysterious metaphysical phenomenon, but just the name we place on what is essentially a very complex mechanical process.

Turing's claim was that if a machine could effectively simulate thinking to the point of indistinguishability from an actual human being, there would be no reason to deny that the machine was thinking. In the larger philosophical picture, this claim belongs to a school of similar claims by empirical philosophers to the general effect that the real existence of [x] is equal to its empirical footprint.

The test is often misunderstood as claiming that passing the test is evidence of intelligence. This is not true to Turing. His claim is that passing the test is intelligence. In other words, what we view as human intelligence is nothing more than passing the empirical test (manifesting the perceivable signs of the patterns we call intelligence). There's nothing else behind the curtain, even for people.

3

You have asked two questions in effect:

  1. To what extent is the Turing test applicable to robots?
  2. How can we compare processing methods of humans and computers?

Question 1

The Turing Test argues that if a person cannot tell the difference between a human and a machine through the course of a interaction then the man and the machine are of comparable intelligence, i.e. the machine is as intelligent as the human. Originally this test was to determine whether a program, a software agent, had achieved human level intelligence. From one perspective it attempts to avoid defining the term "intelligence". Arguably it defines intelligence as equality or similarity in behavior.

Variations on the Turing test have been proposed to address limitations of the original approach. AIMA has examples if I recall correctly but I don't remember any addressing robots. To apply it to a robot would either require an android (a robot that looks like a human), and a very convincing one at that, or the human-robot interactions would have to be limited such that the human cannot see the robot. So long as one accepts the premise of the test, i.e. behavior equals intelligence, then the test is valid up to the outlined limitations.

Question 2

The argument you are reciting i.e. that a machine cannot understand, is called Searle's Chinese Room. Effectively Searle argues that the man in the room, the processor by analogy, cannot understand Chinese if, given some input Chinese phrase, he can algorithmically generate an appropriate response Chinese phrase. There are numerous counter arguments to this claim. One of particular interest observes that the man in the room would not be the center of understanding but in fact the room would be. In other words the room itself is what understands Chinese, the man is merely a part of the system. This is analogous to saying that it is not our brain that understands Chinese but our mind. A subtle distinction but an important one none-the-less. In short, it's unlikely that the computer "understands" in binary. While all input is converted to and processed in binary this is all only part of the process of understanding.

The methods humans use to process input and how it generates behavior is generally not well understood. In theory memory is a function of the weights between neurons and the processing of data is achieved through the interaction of neurons firing. All methods of sensing in the human body reduce to neurons firing which is itself a conversion of data. The encoding we use may not be binary but it is an encoding none-the-less. As such the argument goes that both computers and humans are processing encoded data which brings us back to Turing. We can compare humans and machines by way of the universal Turing machine. If we can show that they are both in fact universal Turing machines then they are both in fact capable of the same things including controversial ideas like being intelligent.

  • I believe that the level of interaction with the test subject could be limited yet still be a valid Turing test. For example, say you were chatting with an entity online. Would you know that it is a human or a machine at the other end just through the context of the conversation? In a way, this limits the test to a judgement of intelligence alone (or whatever you may call it) and not how it appears or moves or inflects its mechanical voice (for example). Seems like a legit test to me. Of course, there are many possible variations. – Octopus Apr 26 '13 at 3:18
  • As I said the test was originally intended for software agents, i.e. programs. Adaptations were developed to test for intelligence through the course of interactions that required more than just chatting. – DaemonMaker Apr 26 '13 at 3:28
3

The most recent article that I've read on this debate is called "Why Philosophers Should Care About Computational Complexity" (PDF) by Scott Aaronson. The discussion of the Turing Test is in section 4 of this paper, but the entire text is very worth reading.

I'll try to summarize it. First, the Turing test is really 2 questions rolled into one:

  1. Does a computer have "consciousness"?
  2. Could we write a program that could pass the Turing test (and if not, why)?

His answer to the second question is to reason that since the Turing test involves a finite number of questions, it's technically possible to just use a huge lookup table -- massively finite, but still finite -- to pass the test. For each question from the human, just look up the response that will work. There are a finite number of lookup tables (again, massive beyond any computing resources that exist today, but still finite), and even though the vast majority of them will produce nonsense, that doesn't change the fact that there exist some number of lookup tables that could each pass the test.

So is passing the the test sufficient for us to say that a computer program has consciousness? That would depend on whether you agree with John Searle (also mentioned in this paper) who has said that if his best friend turned out to be controlled by a computer instead of a brain, then he would consider his friend to have never been a person at all.

So at the end of the day, your job is to define consciousness before you can define a test for consciousness.

  • It was very difficult to resist the urge to make a joke involving Twitter. – Ian Apr 28 '13 at 19:36
  • I wanted to see the joke :( But really, how can one tell if we are not lookup tables ourselves? Well, lookup tables with a degree of randomness (to account for creativity) and maybe a bit of logic (but nothing out of a computer's reach). – Shahbaz Apr 29 '13 at 8:32
  • 1
    The question you're asking brings in a lot of philosophy that I'm not sure I'm qualified to work with. One point I remember hearing is that every choice we make is a reaction to a set of external stimuli (some of it in the form of accumulated knowledge from past stimuli, some of it from current stimuli), so it's certainly possible that we are all some form of lookup table. And we're back to the same question of "can we test wether computers are intelligent without defining how & why we ourselves are intelligent?" – Ian Apr 29 '13 at 18:36
2

The Turing test enables the examiners to see if the judges are unable to distinguish the people from the robots. If the judges cannot tell the difference, then the computer is intelligent enough to appear to be a person.

True, computers convert information, but what if we do as well with neurotransmitters and electrical impulses? What if we cannot truly think, or what if computers can think? Thought is only relative.

Here is an interesting article on thought: http://www.psychologytoday.com/blog/consciousness-and-the-brain/201202/what-is-thought-0

Great questions though.

2

The Turing Test tests if a system can mimic a human being's behavior, not it's inner state. Therefore, what the system does internally is irrelevant.

What does it mean for a system to mimic the inner behavior of a human? Two possibilities come to mind.

  1. The system's internal architecture is similar to that of a human (some work has been done on this -- see Neural Networks) and may include embodiment (see this experiment).

  2. The system has subjective states -- see Artificial Consciousness. This requires knowing if a system is conscious, which would likely require solving the Hard Problem of Consciousness. It even be impossible to know this (see the Philosophical Zombie).

1

As an advocate of the "hard A.I." thesis I'd say that the Turing test is neither necessary nor sufficient for confirming that a robot is intelligent, but that it is sufficient for confirming that a robot has human intelligence.

Passing the Turing test is not a necessary condition for confirming that a robot is intelligent in general, since I can easily imagine aliens who are intelligent but would not pass the Turing test because they are quite different from humans also in their thinking. Hence, it is also not sufficient.

Passing the Turing test is a sufficient condition for confirming that a robot has human intelligence, because we use the same test everyday to confirm that fellow humans are intelligent - insofar as they are.

Regarding consciousness: This could simply be a property of a certain class of computational states of a computer (which is an open system). The fact that we have not yet found out which states and the fact that we do not seem to have an 'intuitive' insight how this could be possible are both irrelevant. If a certain machine was conscious, a skeptic might never know for sure, by the same token as we cannot know for sure that other people are conscious.

1

How do you know other humans are intelligent? You interact with them and know how they behave. If you look into their head with scientific experiments, all you find is movement of atoms and electrons. Now when you look into a computer, it's only electrons that move (well, except for electromigration, which is actually an unwanted effect). But it it really complete atoms moving around that makes intelligence?

Now you may believe that there is an immaterial mind controlling the brain. However even if it is true, we don't know how the mind does that, and we can say for sure that a robot that we would not be able to distinguish from a human would be sufficiently complicated that we could not follow all its processes in detail either, so how can we exclude that a non-physical mind also controls the robot? We don't have a theory of non-physical minds, so there's nothing telling you such a mind cannot be attached to any sufficiently complex structure (indeed, we can't even exclude that a mind is attached to something as simple as a stone; maybe stones have a mind, but that mind is unable to control the stone because the stone is simply not complex enough to allow for that type of control?)

Also note that a robot that is indistinguishable from a human will insist to have a mind and to be intelligent (otherwise we would be able to distinguish it on that difference).

  • The problem with the Turing test is that it relies on convincing people that the computer is human. It is close to a magic show where the magician tries to convince the audience he has real powers. One needs a test that one can tell more objectively if there is intelligence present. Your points about non-physical minds raise important questions. Any theory about them needs to show what has a mind controlling it and what doesn't. The same would apply to strong AI. One of Searle's objections to it was that everything is mental including thermostats. – Frank Hubeny Jan 3 '18 at 16:15

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.