0

Could anyone explain this problem I have with the Turing test as performed? Turing (1950) in describing the test says the computer takes the part of the man then plays the game as when played between a man and a woman. In the game, the man and the woman communicate with the hidden judge by text alone (Turing recommends using teleprinters). If the computer takes the part of the man, then it will have an eye and a finger in order to use the teleprinter as the man would have done. But in the TT as performed, the machine is not robotic. It has no eyes and no fingers but rather is wired directly into the judge's terminal. The only thing the machine gets from the judge is what flows down the wire. But the problem I have is, what flows down the wire is not text. The human contestant gets the text. The judge's questions print out on the teleprinter paper roll. The man sees the shapes of the text, and understands the meanings of the shapes. But the computer is never exposed to the shapes of the questions, so how could it possibly know what they mean?

I've never seen anyone raise this problem, so I'm very confused. How could the machine possibly know the judge's questions if it is never exposed to the shapes of the text?

Roddus
  • 161
  • 6
  • 1
    Rather than writing "Problem with the Turing Test as Performed", can you just put your **specific** question in the title? Thanks. – nbro Jan 31 '22 at 09:04
  • Regarding: "the computer is never exposed to the shapes of the questions, so how could it possibly know what they mean?" could you please clarify. OCR software doesn't "understand" the characters it reads. It may have a language model, similar to the AI's one, to help it disambiguate e.g. `S` from `5`, but it doesn't have any representation of what objects it may convert text for actually *are*. This is a difficult problem in all AI to date, but it is not clear whether you are asking about this. Your description of a physical typing robot would have the same issue. – Neil Slater Jan 31 '22 at 13:58
  • @Neil Slater, so computers internally manipulate instances of binary difference. Text has meaning because a community has assigned an interpretation to the shapes of the text. But no one has assigned meanings to the shapes of the (individual or groups of) binary difference computers internally manipulate (semiconductor switch states and clocked groups of electrons). So what the machine gets from the judge's keyboard and internally manipulate have no meaning. The machine could never understand the judge's questions. But the TT is a test of whether the contestant understands the questions. – Roddus Feb 01 '22 at 21:44
  • @Neil Slater (cont), so for the robot contestant, it "sees" the judge's text then types answers on a keyboard, as a human does. So what travels from the robot "eyes" and to the robot fingers, is electrical pulses (the shapes of which have not been assigned meanings and hence are not text). But then the shapes of the electrical pulses which travel along the optic nerve to the organic brain haven't been assigned meanings either. Yet we understand things. So the key is understanding how knowledge is acquired then accessed, when the things that propagate from the senses have no meanings. – Roddus Feb 01 '22 at 21:56
  • @Neil Slater (cont2), so for OCR, I think that what happens is a direct prescription of human intelligence. So it's the old Ada Lovelace objection that computers (as per Babbage) can only do what they are designed to do (Note G). The human specifies the language model. I think the thing here is to understand how the machine could learn a language without having a pre-existing language model, and perhaps even without having a universal grammar. – Roddus Feb 01 '22 at 22:17

2 Answers2

1

(As @nbro writes, your question is not very specific; I'm answering here how I understand it from the current version)

In an ideal world, a computer would see written text (via a camera), scan it, understand it, and type a response. I assume Turing didn't go for voice transmission, as voice includes other clues to a person's gender.

However, AI is such a complex field that it would have been impractical to implement this until fairly recently. And OCR and robotic movements (typing on a keyboard) are arguably not that relevant to human cognition, so in most actually run Turing-like tests shortcuts are taken.

Update: Also, note that the original Turing test (1950) was based on a party game about distinguishing between a man and a woman (who were not visible). This imitation game was later generalised to a guessing game between a human and a machine.

Oliver Mason
  • 5,322
  • 12
  • 32
  • It's not true that the original Turing test was _just_ about differentiating between a man and woman. It's true that this is how he initially introduces what he calls the _imitation game_, though. Turing actually writes "_We now ask the question, ‘What will happen when a machine takes the part of A in this game?’ Will the interrogator decide wrongly as often when the game is played like this as he does when the game is played between a man and a woman? These questions replace our original, ‘Can machines think?’_". So, the 1950 paper was really investigating the question: "Can machines think?". – nbro Jan 31 '22 at 18:45
  • @nbro Thanks for pointing this out; it's been a while since I read that... I clarified this in my answer. – Oliver Mason Jan 31 '22 at 20:45
  • @nbro, Thanks. Yes, basically. But Turing did reject the idea of thinking as too meaningless to deserve discussion. But as a test of (human-like conversational) intelligence, what is the test for, exactly, and what is the evidence which determines pass or fail? The TT tests whether the contestant understand the meanings of the judge's questions. These questions, being text, are made of shapes. The evidence of understanding is that the shapes of the contestant's text answers, as interpreted by the judge, seem human-like. The TT is a test of understanding of the meanings of shapes. – Roddus Feb 01 '22 at 22:46
  • @nbro (cont), the computer (as electronics engineers know) doesn't receive and/or manipulate text. So if the only thing the machine gets from the judge is electronic pluses from the judge's keyboard, then the machine doesn't get the text of the judge's questions. It has no shapes in respect of which to interpret meanings. Even if the machine could interpret the meanings of text, it could never know the judge's questions because it never gets the judge's text. Since the TT tests for understanding of the judge's text, the test as performed can't possibly yield a result. – Roddus Feb 01 '22 at 23:02
  • @Oliver Mason, Comments I've made above might answer some of your comments. But I think the TT isn't really a guessing game. The judge must rely on evidence. The problem I'm trying to raise is that once this evidence is clearly understood - from the correct perspective of interpretation of shape - fairly severe problems arise. It seems the semantic aspects of the TT as performed have been largely ignored. But the really key issue is understanding how a computer could think. Imagine the huge help true AGI would be in mitigating climate change. – Roddus Feb 01 '22 at 23:23
0

There is no form of OCR that assigns "meaning" by processing visual input of letters and words into the computer representation of those same words (e.g. ASCII). The robot with a camera and keyboard does not solve the problem you have raised. You need to look elsewhere for answers, and the state of AI today is that no-one has strong evidence for how meaning arises within an intelligent system. There is plenty of writing on the subject though.

I think you are trying to understand how and where meaning may arise in any system (biological or machine). There is lots of thought around this subject in AI philosophy and research. A good place to start might be with John Searle's Chinese Room argument, which broadly agrees with you that a basic discussion/chatbot program does not prove intelligence, but for different reasons than the "shapes of the text", which is not really an issue at all in my opinion. Searle's argument is by no means the end of the matter, and plenty has been written in rebuttal and support of the argument.

The real issue is the symbol grounding problem, which also applies to visuals of text, or any system of referring one entity by using another entirely different one.

Potentially addressing the grounding problem, there are various philosophies and engineering ideas proposed. These include:

  • Behaviouralism. It is not important what goes on inside an intelligent system, only that its external measurable behaviours are those of an intelligent system. This matches quite closely to the idea of the Turing Test, but many people find this unsatisfying due to personal experience of self-awareness, subjective experience and consciousness. It is in some ways the "shut up and calculate" of AI.

  • Embodiment and multi-modal experience. If an agent can experience the world directly and associate symbols with relevant experiences (the word "cat" with seeing and hearing cats), then it would be intelligent in the same way as we are.

  • Missing components. Humans (and sometimes animals) possess some additional system that cannot be replicated by current computing and robotic devices, even if they were made 1000s of times more powerful. The missing component might be something quantum in our cells (Penrose, The Emperor's New Mind) or "the soul". This is also a common depiction of robots and AI in science fiction, and there is lots of popular support for it as a philosophy, despite weak evidence.

  • Complexity and power. We can currently replicate the mental power of small insects on computing devices. When we scale up with more powerful computers, larger neural networks, and perhaps a bit of special extra structure (that we don't know yet), then we will hit a level of complexity where true intelligence will emerge. You could view recent very large language models such as GPT-3 as exploring this idea.

Neil Slater
  • 28,678
  • 3
  • 38
  • 60
  • But ASCII isn't a computer representation of text words, but rather a standard table printed on sheets of paper which associates mainly letters of the English alphabet with what are called 8-bit strings. So the symbol "A" is associated with the symbol sequence "01000001". But there are no 0s and 1s manipulated inside computers, so "01000001" can't be a computer representation of "A". I know everyone says computers process zeros and ones, but this simply isn't true. It gives the false impression that computers process symbols when actually they process electrical pulses and switch states. – Roddus Feb 03 '22 at 04:46
  • I agree a robot contestant merely having eyes and fingers doesn't solve the problem of how a system understands the meanings of text. But if a computer will pass the Turing Test, it will have eyes and fingers. So being robotic to that extent is just a precondition. I agree there's no consensus of what intelligence amounts to. But I think we can agree that intelligent systems contain knowledge. So how does the knowledge get inside the system? Through sensory apparatus. So sensory streams contain knowledge in some form. What is the form? Working this out seems a sensible place to start. – Roddus Feb 03 '22 at 04:56
  • CRA. I think basically the CRA shows that no system which manipulates symbols and does nothing else could possibly understand the meanings of what it manipulates (Searle, NY Review of Books, 4 October 2014). He argues that computers are such systems. Hence, they can't possibly understand what they manipulate because, in themselves, those things have no meanings. Searle always comes back to this point. I know there are objections to the CRA, but I've never seen an objection to Searle's fundamental premise that symbols in themselves are meaningless. (But do computers manipulate symbols?) – Roddus Feb 03 '22 at 05:21
  • Symbol Grounding. I distantly recall Harnad's Symbol Grounding Problem in response to which he proposes (something like) analogue processing lower down and symbolic processing higher up. Of course, I've done only only half this - proposed a problem but not offered a solution. The main problem I see with behaviorism is that in effect it's a set of conditionals of the form: If the input = A then the output = B. And B is the response to A not because the machine has applied any intelligence. Rather, the response is a function purely of the programmer's intelligence. – Roddus Feb 03 '22 at 05:33
  • Embodiment and multi-modal experience. You say "If an agent can experience the world directly and associate symbols with relevant experiences (the word "cat" with seeing and hearing cats), then it would be intelligent in the same way as we are". However, we don't associate symbols with experiences. We experience a cat and we experience a symbol. We see both. Both are external. There's no symbol "cat" inside the brain to be associated with an experience of a cat. Similarly, there are no symbols inside computers. – Roddus Feb 03 '22 at 05:40
  • Components of human intelligence that cannot be reproduced in computers (eg soul). Presumably a computer with human-like intelligence could get religion then embrace the belief that it has a soul. But what if there are such things as (immaterial, eternal) souls? But I don't see why an omnipotent omniscient god couldn't create a soul with the owner being a robot, just as a god creates a soul with a human embryo being the owner. For quantum aspects, quantum seems just a current speculation. – Roddus Feb 03 '22 at 05:52
  • Emergent intelligence. All we need is more computer power, processors, memory, etc. This idea has been used since the 1950s to explain failure to achieve milestones. I think if the idea were true, we would have seen successful incremental steps towards AGI. Milestones would have been reached. But the present systems are far removed form anything like human intelligence. The latest approach, Deep learning, has the sever problems of edge cases, catastrophic forgetting, and adversarial attack, which are debilitating. – Roddus Feb 03 '22 at 06:01
  • @Roddus: This is not a discussion forum, and I cannot engage in debate or discussion with you in the comments. If my answer addressed your question as written, please accept it. If it did not, please explain what is missing or what you do not understand. – Neil Slater Feb 03 '22 at 06:47
  • Thanks for explaining the purpose of ai.se Could you suggest a debate or discussion forum? For ai.se, now I can see the problem I've been having. I thought ai.se was a forum where comments or answers could be challenged. But it's really a forum where the established wisdom is amplified (by the like-minded voting up questions) and any dissent is punished (by voting down questions). I was looking to engage with free thinkers who are prepared to justify their positions in the dialectical style where one can learn through an interchange of ideas. I guess that's just not ai.se. – Roddus Feb 05 '22 at 09:02
  • @Roddus the way to engage with discussion here on contentious subjects would be to write your own answers. Votes may fall in line with current consensus on the subject, but if the site is working well, then differing answers on more philosophical subjects may be equally upvoted. It's not a space for debate though. Sorry I don't know of any and the few I have found in the past for AI have been full of cranks, too much noise for me. – Neil Slater Feb 05 '22 at 12:00
  • Thanks for your good advice. – Roddus Feb 09 '22 at 03:15