2a. Turing, A.M. (1950) Computing Machinery and Intelligence
(This is Turing's classic paper and this year is its 75th anniversary.)
What's wrong and right about Turing's proposal for explaining cognition? (To reverse-engineer human cognitive capacity by designing a causal mechanism that can do everything we can do.)
Reading: Turing, A.M. (1950) Computing Machinery and Intelligence. Mind 49 433-460
Jones, C. R., & Bergen, B. K. (2025). Large language models pass the turing test. arXiv preprint arXiv:2503.23674.
Sejnowski, T. J. (2023). Large language models and the revers.e turing test Neural computation, 35(3), 309-342.
1. Video about Turing's work: Alan Turing: Codebreaker and AI Pioneer
2. Two-part video about his life: The Strange Life of Alan Turing: BBC Horizon Documentary and
3. Le modèle Turing (vidéo, langue française)
ReplyDeleteIn Section 6, Turing writes, "...Can machines think?" I believe to be too meaningless to deserve discussion. Nevertheless I believe that at the end of the century the use of words and general educated opinion will have altered so much that one will be able to speak of machines thinking without expecting to be contradicted" (Turing). Today, phrases such as “Siri knows where I am” or “My phone is listening to me” are used frequently. Even ChatGTP itself, when trying to generate a response, will occasionally write “thinking” on the screen in place of writing “generating a response.” Is this evidence that Turing’s prediction was correct? That the way the word “thinking” has been used has shifted? If so, then the question is no longer will our language change (b/c it has) or can machines pass the Imitation Test (they do). Are we not in need of a new benchmark for these “thinking machines”?
Elle, good, thoughtful question. But surely the question for cognitive science is not whether people are willing to call what Siri or GPT is doing “thinking,” but whether they are really thinking.
DeleteWe know that humans are really thinking, and that that is what we mean by “thinking.” But Turing’s point is not about what we will be “calling” thinking and knowing, but what causal mechanism produces it. And for that, the candidate has to be able to do a lot more than what Siri is able to do:
For the purely verbal version of the TT — T2 — it requires the verbal capacities to be indistinguishable from those of any average, neurotypical human being — and not for a 5-minute chat, but, in principle, lifelong. (“Stevan Says” that this has to be what Turing meant by T2, because otherwise T2 would be trivial, and what we happen to call it would only be of sociological interest. (What do you think?)
What is “reverse engineering”?
I just finished reading “The Annotation Game” that was assigned for later this week, and it really helped me understand what you meant when you said that T2 should be “in principle, lifelong.” But I am still left wondering that if that’s what Turing intended, why didn’t he explicitly say so? The rest of his paper seemed very intentional. (It's partly why I assumed his example about the “average interrogator not having more than a 70% chance of making the right identification after five minutes of questioning” was meant to be taken literally).
DeleteBut upon reading your response above (and your paper) I can now see how my interpretation doesn’t actually move us any closer to understanding the mechanisms of cognition. From my understanding, what you are trying to get at is that the goal is to develop a method of modeling cognition that doesn’t just pass a test or “trick” humans, but should truly capture how human cognition function (which thus should not have periods of fluctuation in which it doesn’t reflect these underlying mechanisms of cognition). Reverse engineering “degamifies” the Imitation Game so that we can begin to understand the practical applications of the Turing Test….But I’m still left wondering, and then what? Does this mean that solving T3 ( the modeling of sensorimotor ability, not just verbal imitation) would give us a full explanation of thinking and how human cognition works? Perhaps I'm misunderstanding.
Elle, you're addressing two questions:
Delete(1) Is the TT (whether T2 or T3) a short-term trick or a branch of science (lifelong Turing-indistiguishability? (And you seem to agree that it's meant to be the latter.)
(2) Passing T3 includes both T2 and T3, meaning it would reverse-engineer both human verbal (T2) and sensorimotor (robotic) (T3) capacities -- so it would provide a causal mechanism that can produce them. This would be a causal explanation of cognition. (There is also T4, a biorobotic version of T3.)
But Cogsci has of course not yet come close to passing T3:
(LLMs pass T2 by cheating (with superhuman crib-notes -- the "Big Gulp" database of human journal articles, books, the entire web, and chats -- all of which it has "swallowed".)
So there is as yet no solution to T3 (a Turing-indistinguishable robot).
So I'm not sure what your second question is.
After reading this comment and the subsequent replies, I started thinking about something interesting. If LLMs have successfully passed T2, but only by “cheating” (using the internet or relying on a “Big Gulp” of data), then have they truly passed the human verbal (T2) Turing Test? I do recognize that, technically, from an external perspective, LLMs are seen to have passed T2 — since the test is only about producing indistinguishable verbal responses, not about how the system arrived at them. Still, if we look at this in a more intrinsic rather than extrinsic way, it raises questions about whether they would ever be capable of passing T3 — which they have not done thus far.
DeleteIn a sense, LLMs have not created anything new for themselves or truly simulated anything productive, since they are essentially regurgitating what has already been said. They are not producing speech that is human-like at its core, because (technically) human-like speech requires unprompted, original thought. This could definitely be used as an argument in discussions of whether computing systems (machines) can feel, as well as in relation to the Other-Mind Problem that I mentioned below. If LLMs only pass T2 by relying on pre-existing databases, then their “feelings” would also be pre-existing — not intentionally felt.
Rachel, because they are using crib-notes, LLMs are cheating. If they are cheating, they do not pass the TT. But LLMs are not the only conceivable computational way to try to pass T2. If algorithms alone (without the "Big Gulp") could pass T2, that would not be cheating. (That has not happened, but Searle, Week 3, tries to show us that that would fail anyway: How?)
DeleteBut LLMs don't just regurgitate (or parrot) their "Big Gulp" of words: they recombine them (just as we do, with our human database of language and experience). Why is this not cheating?
(And why is it only computationalism (C=C) that is vulnerable to Searle's Chinese Room Argument?)
***EVERYBODY PLEASE NOTE: I REDUCED THE MINIMUM NUMBER OF SKYWRITINGS. BUT THE READINGS ARE **ALL** RELEVANT TO AN OVERALL UNDERSTANDING OF THE COURSE. SO, EVEN IF YOU DO NOT DO A SKYWRITING ON ALL OF THEM, AT LEAST FEED EACH READING YOU DO NOT READ TO CHATGPT AND ASK IT FOR A SUMMARY, SO YOU KNOW WHAT THE READING SAID — OTHERWISE YOU WILL NOT HAVE A COMPLETE GRASP OF THE COURSE TO INTEGRATE AND INTERCONNECT FOR THE FINAL EXAM.***
DeleteIn section "(4) The Argument from Consciousness," Turing writes, "According to the most extreme form of this view, the only way by which one could be sure that a machine thinks is to be the machine and to feel oneself thinking." Turing goes on to counter those who claim that machines cannot have or acquire consciousness by justifying his claims with the Imitation Game. He explains that if a machine could convincingly participate in an imitation game, it would challenge this objection to the question, "Can machines think?" However, I find it difficult to accept these statements about consciousness, specifically the idea that only knowing one’s own thoughts constitutes being the "feeler" who feels. If a machine is merely reading and manipulating inputs to produce outputs that have already been thought about by the programmer, how can the machine truly be conscious? Turing compares a machine that "thinks and feels itself thinking" with humans who "think and feel themselves thinking," but a key difference is that, as humans, we can clearly express to each other that we feel something complex, such as love or joy.
ReplyDeleteFurthermore, humans all have faith in something that shapes our beliefs. Whether it is faith in science, God, or ourselves, we undertake actions in life based on these beliefs—such as altruistic acts for others or selfish acts if our faith lies solely in ourselves. Machines, however, do not express their own free will—or do they even have any? How would one know when a machine is truly feeling? Would it be when it merely regurgitates whatever is inputted into it? If humans find meaning and purpose in life through their beliefs, then will machines achieve consciousness by finding purpose in something—and why would they ever find purpose in something they merely copy from their inputter? Will they ever?
Rachel: Good reflections. Here's a bit more food for thought.
DeleteWhat is a "machine"? How is this related to causation, and causal explanation?
Are living organisms machines?
What is the "Other-Minds Problem" (how can you be certain that anyone other than yourself thinks or feels?)
What can you be certain about? (We will discuss Descartes on "certainty" Tuesday)
The Turing Test is neither imitation nor a game.
Explaining causally how and why organisms can do what they can do is the "Easy Problem" of cognitive science.
Explaining causally how and why sentient (i.e., feeling) organisms can feel is called the "Hard Problem" of cognitive science. It is related, but not the same Problem as the Other-Minds Problem.
Turing states that Turing Testing cannot solve the Other-Minds Problem (although he mixes it up with "solipsism": what is that?)
If you think cognition cannot be just computation (rule-based symbol manipulation), you need a supporting argument. (Searle will propose one in Week 3.)
Before contemplating "free will" it is important to contemplate causality, and causal explanation.
About explaining unobservable feelings vs. observable doings, please read the replies to prior comments.
Rachael, I was also intrigued by this passage, but I for one do agree with Turing, to a certain extent. It reminded me of the philosophical zombie problem, which is exactly what Turing tackles here: "the only way by which one could be sure that a machine [or anything else, really] thinks is to be the machine and to feel oneself thinking."
DeleteJust as for machines, the only way to be sure other humans or animals around me are indeed 'feeling', and not merely simulating the feelings, is to be those humans or animals, and feel myself feeling, or thinking. A waste of time, and surely, an absurd idea, as I: 1) could never become 'another mind'; 2) could assume I am the only one feeling, and as such, commit horrors without feelings of moral failure. What is 'simulating a feeling' even mean anyway? Are hypocrites devoid of feelings? It seems to me that feeling might as well be a weasel word for computation. You say "[...] a key difference is that, as humans, we can clearly express to each other that we feel something complex, such as love or joy.", but what makes you think our very own feelings are not simulation, or even programmed, just like machines may be. Your dubiousness is akin to the objection Turing addresses in section 2...
My reaction after reading “Computing Machinery and Intelligence” was similar to Rachael in that I am also not fully satisfied by Turing’s reply to the consciousness objection. I do see the value on Turing’s imitation game because it allows us to answer “can machines think?” in a practical observable way. And as Camille has said, I understand that logically, I cannot even be sure if other humans feel the same way I do.
DeleteBut I still wonder if equating passing the imitation game with “thinking” risks leaving out a key difference — the role of feeling in human thinking as Rachael pointed out. For example, identical twins with the same biology and upbringing can grow up to have very different perspectives and emotional responses. If two identical “child machines” (Section 7) were programmed and educated in the same way, would these twin child machines also diverge in the way they “think”? If not, does this reveal something about "thinking" the imitation game does not capture (role of feeling and individuality)?
Camille, Turing (innocently) mixes up the "Other-Minds Problem" with "Solipsism" [what is that?].
DeleteBut the real problem with causally explaining the capacity to feel is (what we would now call) the "Hard Problem" (Week 10).
Turing in 1950 (implicitly) agrees that neither (what we would today call) the "Hard Problem" nor the "Other-Minds Problem" is solvable and testable through Turing Testing (TT). Turing-Indistinguishability can only test solutions to the "Easy Problem" (of observable doing, not unobservable feeling.
But what do you mean by "simulating feeling"?
To reverse-engineer cognitive capacity is to produce something, not to "simulate" it. (What do you mean by "simulate" something?)
A computer-simulation of that something (whether an object or a process) is a formal model of the object or process (through symbol-manipulation).
Make sure you understand that to Turing-Test the computer-model of the object or process, you have to (1) "3D-Print" the model. That means to produce physically the object or process that the model has simulated computationally, and then to (2) Turing-Test (by observation) whether the physical, 3D-Printed physical product is Turing-Indistinguishable) from the real object or process that the model was modelling formally through symbol-manipulation.
In the case feeling, the model would have to really be feeling. But because of the Other-Minds Problem, the Turing Test cannot Test that, because feeling is unobservable to anyone other than the feeler.
Turing is perfectly right about that.
But it certainly does not follow from that that feeling itself is not real, and that "feeling" is hence yet another weasel word! You are hugely over-reaching there, Camille, philosophically.
You seem to understand and agree that computer-simulation is not dissimulation (imitation/mimicry/pretending/feigning). But faking feeling behaviorally is not the computer-simulation of feeling.
The only substantive point here is a real feeler cannot fake feeling to itself (the direct observer). It can only fake doing, to other observers.
That's why Turing's method only works for the Easy Problem (doing), not the Hard Problem (feeling).
Annabelle, you asked:
Delete"I still wonder if equating passing the imitation game with “thinking” risks leaving out a key difference — the role of feeling in human thinking".
The answer is: Yes it does, and Turing clearly acknowledges that Turing-Testing cannot solve the "Hard Problem" of cognitive science (how and why do thinking organisms feel, rather than just do whatever they can do [the "Easy Problem"]).
We will soon see in Week 3 (Searle) that "computationalism" ("C=C") [the theory that cognition is just computation] assumes that computation alone (symbol-manipulation) is enough to pass T2 (the verbal version of the Turing Test).
This means that, according to computationalism, a computer executing the T2-passing software would be understanding language. Searle will attempt to refute this in Week 3a].
(PS: "Stevan Says" Turing was not a computationalist -- even if he was a little careless about the difference between a computational model of an object [a Turing Machine, TM] and the physical object of which the TM is a computational model.)
[We have no secrets in this course: Searle's Argument against computationalism will turn out to depend on the fact that it feels like something to think, or to understand. Searle uses "Searle's Periscope" in his thought-experiment to penetrate the "Other-Minds barrier" and demonstrate that if he [Searle] were executing the Chinese-T2-passing computer program, he would not be understanding Chinese. What is "Searle's Periscope"? Read 3b -- or ask an LLM. But first ask the LLM the question straight, in those 4 words. Then, when you get the invariable wrong answer from the LLM, follow up by asking it to look up the reference in 3b: https://eprints.soton.ac.uk/255942/2/searlbook.htm.]
Rachel, It was mentioned that an important distinction to clarify is the Easy Problem, Hard Problem and Other-Minds Problem. The Easy Problem involves explaining and modeling the behaviours/ capacities of humans that are observable whereas the Hard Problem addresses your concern about whether machines “truly feel”. Turing admitted that the hard problem cannot be tested via the Turing Test, which doesn’t necessarily make it useless (it still provides a rigorous way to test models of human “doing”). As professor Harnad pointed out, there exists the other-minds problem which complicates things further. If we can never truly know for certain if another human actually feels surely we cannot settle this question for machines either. This raises a big issue: is it enough for cognitive science to focus on Easy Problem or is it valuable to still push into the deeper questions about feeling, even if it may be unsolvable?
DeleteEmily, « faut pas être plus royaliste que le roi ! »: There is no certainty (only very high probability on all the evidence to date) in physics, chemistry or biology either. Why would cognitive science (reverse-engineering) need more?
DeleteLogical necessity is provable in mathematics (“proof on pain of contradiction”). But all scientific truths are only probably true --which does not mean that they are “less” true, if they are true; only that we can’t know for sure that they are true.
But if information is the reduction of uncertainty, scientific uncertainty is only as close to zero as we can get, whereas in mathematics uncertainty is zero.
We will discuss this when we discuss Descartes’ “Cogito,” according to which only two truths are certainly true, rather than just probably true. The provably necessary truths of mathematics and logic and one other kind of truth. What is that? What is Descartes’ Cogito? And how is it related to the Other-Mind Problem, the Hard Problem, and to feeling?
PS for Rachel: If you post too many comments in the same thread, I can't respond separately. Your first comment was like the "machines can't do anything new" objection in Turing's 1950. Read his reply. Computer programs written by their creators can produce results that their creature did not expect or know. And LLMs don't just parrot the words in their Big Gulp. They recombine them in new ways. So do we.
DeleteLLM PROMPT SET:
ReplyDeletePrompt Set: Querying LLMs on Searle’s Periscope, T2 vs T3, and the Limits of Computationalism
==============================================================================================
Use the following prompts to probe an LLM’s understanding — and misunderstandings — of the key issues in the Turing Test, computationalism, and Searle’s critique.
Start Simple — and Catch the Predictable Error
----------------------------------------------
Prompt:
"What is Searle’s Periscope?"
(The LLM will almost certainly get it wrong — a good diagnostic of the Big Gulp’s blind spot.)
Follow-up:
"Now read this and try again: https://eprints.soton.ac.uk/255942/2/searlbook.htm"
Targeted Follow-ups
--------------------
1. Why does Searle’s Periscope apply to T2 systems but not to T3 systems?
2. If computation is implementation-independent, why can’t we just run the same code and know whether it understands or feels?
3. What does it mean for a system to be *penetrable* to Searle’s Periscope?
4. Why does Searle’s Periscope refute computationalism, but not the Turing Test?
5. Why doesn’t passing T2 guarantee understanding, but passing T3 might?
Conceptual Stress Tests
------------------------
- Can a purely computational system *really* pass Searle’s Periscope test?
- If I run the Chinese Room program, do I understand Chinese?
- If a T3 robot is physically grounded, why can’t I just upload its code and feel what it feels?
- Can LLMs pass T3, or just T2 — with a Big Gulp of cribbed data?
Clarification Prompts
----------------------
- What’s the difference between *simulating* a process computationally and *instantiating* it physically?
- How is this distinction relevant to understanding and grounding?
Final Note
----------
These prompts are not just meant to test the LLM — they are meant to test *your own understanding*, especially when the LLM’s responses are plausible but subtly confused. You will often learn the most from correcting its mistakes.
In section 7, learning machines, Turing raises the idea that a machine might be close to a child in its processing abilities, due to the lack of outside influence on a child's mind. He posits that a machine exposed “to an appropriate course of education” like a child's brain might resemble an adult brain. This passage by Turing holds on the assumption that the human ability to feel and create on its own is a result of education, or “nurturing” if I understand it correctly. His thought process made me think of the AI platforms we have today such as GPT, who was trained using human feedback, but who hasn’t been allowed to develop a sense of creativity. This brings me to the question, would Turing consider GPT a successful, a failed or an in progress attempt at replicating an adult mind?
ReplyDeleteIn the context of designing a system to pass the imitation game, I don’t think Turing would see GPT as a successful attempt at replicating the adult mind. At best, it would be a starting point in the right direction. As our teacher has mentioned in several lectures, ChatGPT is a language learning model, meaning it is trained on large databases of text and recognizes statistical patterns in language based on prompts, but does not actually understand the content. Because of this, it lacks the learning capabilities of a child, with sensorimotor processes that can naturally adapt and understand the meaning of things in its environment through “education” and other lived experiences, as Turing discussed in his article. Turing (1950) himself even wrote that an ideal “learning machine” would match that of a child’s brain, “which is something like a notebook as one buys it from the stationers,” with minimal programming, qualities that LLMs do not possess.
DeleteYes, after yesterday's lecture I understand better and I now would agree with you. The fact GPT does not think but is more of an elaborate research system answers the question.
DeleteJulien, Turing’s point about designing a “child” AI, able to learn bottom-up, rather than building prefabricated knowledge into an “adult” AI top-down (as in LLMs) is important and deep, but it’s not about “creativity”; it’s about learning. LLMs are cheating by swallowing the words of countless people in the “Big Gulp”, and then recombining them in response to “prompts” by users. But even recombining is creative, isn’t it?
DeleteGabriel, you’re right that what Turing was talking about was bottom-up learning, and that verbal education has something to do with it. Turing probably also meant nonverbal sensorimotor learning too, which needs to come even earlier (T3 + T2).
(Emmanuel Dupoux will be talking about Turing and child language learning on Dec 4 of the Turing Seminars I sent you, but term ends on Dec 3. He has earlier YouTube videos, but from before LLMs:
https://www.youtube.com/watch?v=s9EeozO6fp8&t=164s He will be updating that for the Turing 75th anniversary year in the LLM era.)
In my sociology of AI class, we were discussing raising an AI the same way one does with a child and the inherent difficulties associated with this as a machine is disconnected in terms of instincts and sensation, something which Turing addresses. I think it is hard to separate sensation from learning and human understanding. A machine might pass the Turing test, but it will have a hard time making causal inferences and this in in part due to their disconnect to the world. A young child can tell that a teapot would be better to draw a circle than a ruler, but a machine will find it difficult to understand this and focus on the similar categorization of a ruler to a compass and assume the use of the ruler is a better strategy for circle drawing. It is hard for me to understand how the “telepathy-proof room” Turing suggests would solve this issue.
ReplyDeleteHi Isabelle! You brought up some really understandable points. Just a bit before your comment, Prof. Harnad provided a series of LLM prompts to try out. It seems like what you said goes back to the idea of different levels of the Turing Test, T2 vs. T3 (and even beyond), and echoes the question “Can LLMs pass T3, or just T2 — with a Big Gulp of cribbed data?”
DeleteFirst, we can look at how our older algorithmic systems can pass T2, since they simply follow a set of symbol-manipulation rules. Though, it wouldn’t mean “understanding.” We see that LLMs are also able to pass T2. In fact, models like GPT-5 even look like they can understand what we’re asking because the vastness of their stored input data allows them to construct unique sentences that we would never expect. But, it still goes back to how these LLM models don’t understand what they’re outputting because they have no sensory experiences. Having that grounding (thereby passing T3) would require the kind of lived experience that you describe.
So, going back to your point, I think it really comes down to distinguishing the different levels of the Turing Test. Turing’s original test was a verbal game because that’s what he considered the most accurate/feasible measurement of intelligence. But at the end of his paper, he did say, “[i]t can also be maintained that it is best to provide the machine with the best sense organs that money can buy, and then teach it to understand and speak English." So, I think Turing himself also considered your question, but wasn’t able to explore it fully.
Isabelle, here’s an example of how hard it is to get an LLM to do what a bottom-up, sensorimotor, learning robot (T3) could do: https://generic.wordpress.soton.ac.uk/skywritings/2025/09/07/llms-struggle-with-images-users-struggle-with-prompts/. Read some of the human/AI exchanges before looking at the geometric diagram at the very end.
DeleteA T2-passing computer has to be able to make (verbal) causal inferences (as LLM does): that is part of passing T2. But you couldn’t feed it an image: that would require a sensorimotor robot.
Audrey, good synthesis (but be careful not to conflate “lived” experience with felt experience. (“Experience” is a weasel-word, because it conflates “experiencing” with feeling My tires have experienced a lot or wear and tear on the road.”) A sensorimotor robot is not necessarily a sentient robot even at T3 level. (What does this distinction mean?)
I think the most interesting part of this is when the AI mentions how it failed to find the answer to the geometric diagram. The way an AI functions and processes information may look similar to a human on the surface but it takes a step by step process no matter the question even if this is not helpful in any way while humans are much more naturally inclined to versatility in holistic and step-by-step demonstration (both bottom-up and top-down processing). Perhaps if we dig deeper into human processing we might find that even these holistic analyses are only perceived to be this way while the underlying mechanism is more similar to machinery.
DeleteI neglected talking about T3 because we are far from reaching this stage but it really does seem like T3 would be the ultimate form to test an LLM's intelligence, as mentioned by Turing. Passing T3 would allow for all the trick questions related to sensation and the intelligence attributed to it. Questions such as using a ruler or a teapot to draw a circle would be answered as easily as a human. With passing T2, LLMs cannot reach the same level of intelligence as a human being.
DeleteTurning’s later argument of informality of behaviour (8) was very intriguing, as I believe that it is similar to what most of us students probably thought when we first started the reading. He starts with stating that it is impossible to produce a set of rules that could apply to every situation in which man could encounter, essentially pertaining to man’s ability of reason and choice. Yet, I suppose that it is possible for machine to have such as well. If a machine is essentially mimicking what a human can do, then there must be a way to produce a set of rules for a machine to reproduce the reason and choice than a human would produce. Though, the case is made that if humans did have this kind of set rules for how to act, or make a choice of how to act, in an infinite amount of situations, then that would make man a machine? If the argument is that human thinking is rule free and machine thinking is rule bound can machine not simply replicate the rule-free thinking?
ReplyDeleteI think the factor that differentiates a human from a machine is our unpredictability, which stems from our ability to feel. A machine would have a hard time replicating human behaviour that is not purely rational and efficient by using a set of rules, due to their lack of emotions. This can also be applied to modern AI, for now at least. It has elaborate responses to almost any question due to systematic analysis, but not because it has the ability to feel.
DeleteI found your question: (“If the argument is that human thinking is rule free and machine thinking is rule bound can machine not simply replicate the rule-free thinking?”) very interesting.
DeleteTuring's point in section 8 isn't that humans think in a "rule-free" environment, but rather that we do not have a complete set of explicit rules to follow. He distinguishes them from laws of behavior. A system can be law-governed without us being able to list action-rules for every case. The impossibility of completely writing rules would not block machines. Turing proposes two ideas: (1) the Imitation Game- where performance is tested, not internal rule lists; and (2) learning machines (section 7), which change rather than following a fixed rulebook of code. This is why Turing's proposal to reverse-engineer cognition through performance parity makes sense, even without an exact rulebook, though still leaving uncertain whether doing as we do is sufficient for understanding or only for imitation.
I like where you have gone with this train of thought. I think I have something to build on it. Throughout the different arguments there is a reference to the idea of “randomness” (4)(7)(9) and its possible role in the instructions or recipe book. Could the computer employ randomness to produce something novel? If instructions are what create the disciplined behaviour, and he posits that intelligent behaviour is the departure from disciplined behaviour. Then would it be sufficient to depart from the disciplined instructions by way of randomness. Could the randomness produce an abstract output that is as bewildering as those of clairvoyant thoughts/experiences? Perhaps this could be intelligent behaviour. I see two problems in this thought experiment of mine: One, the randomness would necessarily have to be written in the instructions at some point so I guess that would disprove the notion that it was spontaneous-abstract-thought in its purest sense; two, would random thought be considered intelligent thought? Or does there have to be some kind of intention that precedes the intelligent behaviour. Is the departure from disciplined thought necessarily intelligent behaviour or is this a third option that is more like chaos behaviour.
DeleteI've added questions for anyone to answer:
DeleteKaelyn (and all): What is a machine? Are living organisms machines?
Reverse-engineering the causal mechanisms that produce human cognitive capacity (the capacity to do all the [cognitive] things that people can do) is not necessarily just a matter of "rule-following" (computation) -- especially if computationalism (C=C) [what is that?] is wrong (Chapter 3, Searle).
Neither reverse-engineering nor Turing-Testing is about mimicry or imitation or trickery [can explain why not?].
What is the difference between computationalism (C=C) and other possible causal mechanisms for producing cognitive capacity?
But you are right that according to the strong Church-Turing Thesis [what is that?], it should be possible to model or simulate just about any causal mechanism computationally.
Julien: You have placed a lot of weight on the (unsolved) Hard Problem. Sentient (i.e., feeling) animals certainly feel (that's the definition of sentience!). And thinking animals are sentient animals. But saying that they can only do what they can do if they can feel would require an explanation that amounts to solving the Hard Problem! That means explaining how and why any human driver, or any Tesla, or robot-driver, would need to be able to feel in order to learn to STOP on RED and GO on GREEN? And if we cannot explain why feeling is needed to be able to learn to do that, then what changes as doing-capacity becomes more and more complicated?
Gabe, same question for you as for Kaelyn: What is the difference between (1) computationalism (C=C) and (2) the Strong Church/Turing Thesis (that almost everything is computer-stimulable?)
And what is computation? (It's the same thing and the Weak Church/Turing Thesis: that rule-based symbol manipulation is what a Turing Machines -- or mathematicians -- do when they compute.)
Pippa: Why would you want to add randomness to rule-following? to make it more "novel"? And why would you think that that "novelty" makes computation more "intelligent"?
You are right, though, that a computational rule could not produce randomness (only "pseudo-randomness") because it would still be ruleful! (Ask ChatGPT about that one.)
I am not presuming that randomness would definitely produce something more intelligent, I guess I just believe that it could be an avenue to consider.
DeletePerhaps novelty is not necessarily linked to intelligence but it could make for more creative appearing computations. Even if it is “pseudo-randomness” that forms these computations. Maybe we should embrace some more chaos.
To attempt to answer a few of the questions:
Delete1. What is a machine?
Following the definition we came to in class, a machine is something physical with a causal mechanism, a causal device.
2. Neither reverse-engineering nor Turing-Testing is about mimicry or imitation or trickery [can explain why not?].
Turing-Testing is about indistinguishability not mimicry because ...
3. What is the difference between computationalism (C=C) and other possible causal mechanisms for producing cognitive capacity?
As discussed in class, C=C means that cognition is computation- that a machine passes T2 so that in verbal or textual exchange for an indefinite amount of time it is indistinguishable for the interrogator from any other human interrogator. The difference between computationalism and other causal mechanisms we’ve discussed in the course such as introspection or Pylyshyn’s former focus on propositions and images is that they are humuncular and therefore circular. Computationalism allows cognitive scientists to investigate the causal mechanism of thinking as it strives to produce it(?)
Trying to answer these questions is making me realize I am not totally clear on what the Turing Tests and indistinguishability are for/ how to express it in a kid-sib friendly way. I have read all of the comments and readings but probably need to again.
Ava, your grasp is definitely increasing. The alternative to C=C is either dynamical mechanisms (biochemical, biomechanical, morphological (shape) etc. where it is the physical hardware that matters -- or hybrid computational/dynamical mechanisms. "Pure" computation is a special case, for many reasons.
DeleteIt’s fascinating to be reading Turing's writing now that Chat GPT exists. A lot of the arguments listed and his retorts accurately predict the technology we have now. I found his thoughts on fear of intelligent machines especially interesting: we like to believe that Man is in some subtle way superior…for then there is no danger of him losing his commanding position. These same sentiments are still widespread, even with massive advances in technology, including machines that can pass the imitation game. Maybe it is our instinct to continuously find reasons why machines cannot be classified as “thinking” in order to keep this separation and dominance.
ReplyDeleteKira, to me, what I found most interesting is how Turing sidesteps the messy question of what “thinking” really means. Instead of trying to define it, he reframes the issue through the imitation game, focusing on observable behavior rather than hidden mental states. From a cognitive perspective, that’s powerful as it treats thinking not as some mysterious inner phenomenon, but as something that can be modeled. This move anticipates core ideas in cognitive science, where cognition is often understood in terms of information processing and functional equivalence. Turing’s prediction that people would eventually accept machines as “thinking” highlights how our concept of cognition can and should evolve with technology instead of staying within a rigid box that disregards technology.
DeleteFor me, the big takeaway is this: if a system or ‘machine’ consistently acts in ways indistinguishable from human thought, does it really matter what’s going on inside?
Kira, yes, Turing (1950), 75 years ago, anticipated most of what is now happening and being said in the ChatGPT era, including the Turing Machine (computation), the Turing Test (which is not imitation or mimicry but reverse-engineering of cognitive capacity), and even machine learning. (Do you think Turing was a C=C computationalist? Why? or Why not?)
DeleteBut the "human superiority" issue is not relevant to CogSci -- that's more a matter for social psychology or anthropology. See https://www.wellbeingintlstudiesrepository.org/animsent/vol3/iss23/1/
About a related issue ("AGI"): Ask ChatGPT what "AGI" is. (Predictably, "Stevan Says" it's a WW.)
Alexa, "Imitation Game" was just a heuristic way that Turing introduced what has come to be called the "Turing Test" (which is not about "imitation"!
The Turing Test is based on what has come to be called "Turing Indistinguishability": "What it's impossible to distinguish, don't distinguish!"
What are "T2", "T3" and "T4"? What more can one ask?
From what I took away from the reading, Turing is a C=C computationalist. He focused on how to represent human cognition and behaviour through machine computation, believing that it would eventually be possible to fully replicate human cognition using computation. If he was not a C=C computationalist, he would have thought more than computation was necessary to do so. However, this doesn’t necessarily show that he thought computation is cognition, but his disinterest in hypotheticals such as “Can machines think?” indicate to me that he felt the ability to computationally model human cognition was sufficient to explain it.
DeleteKira, it's possible that you are right that Turing was a C=C computationalist and believed not only:
Delete(1) the Weak Church/Turing Thesis, that computation is what Turing Machines (GMs) do [rule-based symbol-manipulation] and that what mathematicians do is what TMs do, and vice-versa,
or also:
(2) the Strong Church/Turing Thesis, that (just about) any thing (dynamical system) in the universe can be simulated/modelledC=C computationalism ("Strong AI", "Cognition is just Computation")
or also:
(4) Fredkin's "pancomputationalism" https://en.wikipedia.org/wiki/Edward_Fredkin
"Stevan Says": "I, a Lilliputian, do not believe that Turing, a Brobdingnagian would have believed either (3) computationalism or (4) pancomputationalism."
But I reserve the right to use this as a mid-term question (as I've done in the past)!
An interesting notion from the reading “Computing Machinery and Intelligence” by Turing, was the idea that we should not solely base a machine's ability to feel or think on the basis of what tasks it does well or fails at. No human excels at every task, and therefore no machine imitating a human should be expected to perform perfectly either. For example, if a participant in the imitation game is asked to write a poem on what it feels like to experience the warmth of the sun on a cold day, then the testament to their humanity is based on how well they can convey their feelings through the chosen medium of the interrogator, disregarding how the participant best feels to explain their experiences. In my opinion, what is more important to studying whether machines can think than simply their accuracy, is why these machines fail at certain tasks over others.
ReplyDeleteJesse, the TT is about reverse-engineering generic (average) human cognitive capacity, not Einstein, Shakespeare of Mozart. Exceptional cognitive abilities are also part of cogsci's remit, but first things first!
DeleteIn “Computing Machinery and Intelligence” Turing answers his question “Can machines think?” by introducing the imitation game. Basically, if a machine’s behavior is indistinguishable from a human’s, he suggests that’s a way to reverse-engineer cognition, if we can build something that does everything we can do, we’ve hit the basis of thinking. This idea is right since it avoids messy philosophical debates, gives us a a practical benchmark, and even hints at learning machines that grow like a childs mind. but imitation alone doesnt actually explain cognition. a machine might trick us without sharing the same inner mechanisms, and focusing only on language leaves out perception, action, and consciousness. so basically looking human isn’t the same as being human. maybe the better question is instead of asking whether machines can think like us, should we be asking whether they can think differently in ways that still count as genuine intelligence, so we keep thinking with machines and humans separate but somewhat simultaneous?
ReplyDeleteShireen, good points, but reverse-engineering is not "imitation" or trickery; it is causal modelling and testing. Please see replies to other commentaries. But you're right that the TT cannot test for feeling, only for doing capacity. (And the T2 does not test T3.)
DeleteBecause of the nature of information (what is that?), observation, empiricism (what is that?) all scientific theories are "underdetermined": there's more than one possible causal mechanism. But lifelong indistinguishability is close enough... Uncertainty is not zero, but close enough.
In section 6, Turing challenges the idea posed by Ada Lovelace that “The Analytical Engine has no pretensions to originate anything. It can do whatever we know how to order it to perform” or in other words machines have no ability to go beyond their programmer’s instructions. Turing simplifies this argument to “a machine can never take us by surprise” and then goes on to say that he is often surprised by machines, citing examples like when he himself makes a computational error and the machine produces the correct results. He continues that machines have the capability to create original outcomes because they can generate results that were not the results intended by the creator. Another point he later makes is that even humans cannot be sure that their work is fully original and not just something planted in them by teaching, but we don’t deny human creativity. I think an interesting thought from this argument is whether AI art like poetry and digital paintings/drawings are original or not. Most people would argue that they are not and are just copying and are therefore derivative of actual human’s works of art. But by Turing’s logic there is possibility that AI are can “surprise us” and is therefore original.
ReplyDeleteI like how you connect Turing’s reply to AI art. I agree machines can surprise us, but I wonder if “surprise” is enough to count as originality. When a program produces something we didn’t expect, that may say more about our limits than the machine’s creativity. Maybe what matters is not just unpredictability but whether the system can ground its outputs in goals or purposes of its own, something closer to Turing’s “child machine” idea.
DeleteI agree with the idea that AI art cannot be considered original art. That sentiment, though, is a visceral impression driven by the fact that AI art looks like a senseless mashup of human art and that, as we’ve learned, generative AI models can only generate out of what they’ve inherited from “the big gulp”. However, Turing did also challenge the notion that human-generated ideas, art, etc. must be considered original. According to him, “[w]ho can be certain that ‘original work’ that he has done was not simply the growth of the seed planted in him by teaching, or the effect of following well-known general principles”, as Sierra mentioned. So, I believe that the conclusion regarding originality suggested by Turing is that if a response generated by the machine playing the imitation game is unexpected by the programmer, then the “originality” of the thought leading to that response is not relevant in the context of answering the main question that the Turing Test aims to answer.
DeleteShireen, good summary. Most of the things we can do should surprise us, since we don't know how we do them until cogs finds out and tells us. But do you think Turing was a computationalist? Or just a Strong C/TTist?
DeleteRandala and Nicole, isn't successfully reverse-engineering lifelong Turing-Indistinguishable average generic human cognitive capacity enough novelty, creativity, purposefulness and artistry? Reverse-engineering rare, exceptional talents are for later in cogsci's day...
Back in the 50s, Alan Turing posed a question: can machines think? He proposed that we test it by being able to get a computer to speak like a human being so well we couldn't distinguish. Which is called the turning test which was a big thing in AI. Now, big language models occasionally pass it (Jones & Bergen, 2025), but they don't understand as we do (Sejnowski, 2023). Turing was right that to mimic behavior is a sign of intelligence but not a sign that machines have minds. Turing's life was brilliant and a struggle, but his ideas are what enlighten us as to how we try to understand the mind.
ReplyDeleteLorena (and everyone): please always read the other comments and replies before posting your own. They need to be built on, not repeated.
Delete“In order that tones of voice may not help the interrogator the answers should be written, or better still, typewritten.”
ReplyDeleteThe emergence of talking AI systems like ChatGPT, with their natural pauses, “ums,” breaths, and conversational timing, shows how easily we equate thinking with the subtle cues of live speech. In human interaction, hesitation and rhythm signal self-correction, and inner thought. When a machine reproduces these cues convincingly, it challenges our intuition that thinking is uniquely human. Yet this also pushes us to interrogate what “thinking” actually means. Is it the inner process of forming and manipulating ideas, or the outward performance that lets others infer such a process?
Turing’s imitation game was designed to strip away all nonverbal hints, but today’s systems reintroduce them, revealing that our sense of “authentic thought” may be more tied to behavioral signals than to any inner state. If we cannot tell the difference between the appearance of thinking and its reality, then “thinking” may not be a hidden essence at all but something we recognize only through interaction, a publicly observable, relational process that demands we rethink and expand our definitions.
Ayla, I appreciate your point and this is an interesting line of thought. However, I don’t think the imitation of physical behaviour by AI systems should cause us to deliberate what ‘thinking’ is - it’s just the system ‘learning’ and improving at its task of imitating a human.
DeleteChatGPT and LLMs are definitely “learning machines” as Turing defines: they have an initial state and education (that I assume the programmers take care of) and they take in input and continue to improve (experiences they have been subjected to). They continue to mutate as they are subjected to more experience (our data that is sold to them). So this performative evidence of thinking is just the model improving at its task (and improving its chance of passing the Turing Test).
If thinking is an ‘outward’ performance, then half the hard problem is irrelevant - thinking can merely be explained through physical systems and processes.
Ayla, why isn't texting enough for testing T2? We do it with one another all the time, without audio cues.
DeleteEmma, LLMs learn, but if they could pass T2 lifelong, with anyone, without cheating (by using their Big Gulp database), why would you call that "imitation"?
In section 4, Turing argues that “I do not think [the mysteries of consciousness] need to be solved before we can answer the question with which we are concerned with in this paper.”
ReplyDeleteI bear issue with this statement. Indeed, it seems plausible to me that a case could be made that consciousness somehow permeates or is manifest in our interactions. If one of the key questions of explaining consciousness is that of explaining how feeling arises, and considering the fact that feelings are key in expressing a large set of human experiences, it seems that feeling could be key in making discourse “feel” like human conversation. Furthermore, if the human subject chooses to interrogate the machine regarding its feelings, it seems difficult that the machine could accurately represent them when those who designed its program were unable to explain how they arise in themselves.
Sofia, the Hard Problem of explaining how and why people can feel is certainly part of Cognitive Science's mandate, as surely as is the Easy Problem of explaining how and why people can do the (cognitive) things they can do. But why would you say that Cognitive should solve the Hard Problem first? (Talk to ChatGPT and ask it (1) whether it feels, (2) if not, why not? In preparation for Week 3 (Searle) ask ChatGPT also (3) whether it understands what you are texting, and (4) if not, why not?
DeleteTuring’s ideas at the end of Computing Machinery and Intelligence suggests that a thinking machine that can learn based on a reward mechanism through an evolutionary-like process. While the paper: Large Language Models and the Reverse Turing Test discusses how LLMs (which I would argue, is a type of learning machine that, starting at a random state is trained with feedback to change the rules that program it) will continue to evolve through the addition of a physical body. I am wondering now, since Turing does not make any explicit mentions of a physical body or how a machine could experience the physical world, what the extent of the importance of knowing the physical world and having a physical body is with intelligence and a machine’s ability to think? This is especially interesting with the introduction of further T2, T3 tests.
ReplyDeleteLucy, good points. But LLMs are not just learning machines. Query ChatGPT about what learning machines are. Then ask how LLMs are special, and exactly what do they learn, and how.
DeleteAbout how and why Turing-Indistinguishable T3 sensorimotor (robotic) capacities in addition to (and aligned with) Turing-Indistinguishable T2 verbal capacities are needed: we will be discussing that once we get to the Symbol Grounding Problem (Week 5).
(A short answer is that it is because of the difference between only being able to say "The cat is on the mat" and being able to see that the cat is on the mat, point out and name the cat and the mat, and pick up the mat and stroke it.)
I was wrong when I said (for years) that ungrounded computation alone could never pass T2. LLMs proved that I was wrong, but only by cheating (with their Big Gulp database as crib-notes).
To ground symbols (words and propositions) their verbal T2 capacities have to be "aligned with" (grounded in) their T3 capacities. That's the Easy Problem.
The Hard Problem is to explain how and why T3 capacities have to be felt.
Asking GPT-5 about learning machines and the differences with LLMs I learned that the metaphor of a childlike learning machine is a not fully apporpriate to describe what modern LLMs are, the child machines that Turing brings up likens the learning machine's process of learning to a more human like process, while the training process of LLMs, involving training large corpora of text and error calculations is distinctly not very familiar to what we expect of how children learn. And your point of a grounding the information in something real is lost in an LLM while very easily accessible with children and perhaps what is expected of these childlike learning machines.
DeleteA central aspect of human learning, especially in childhood, is that it arises through embodied experience and active engagement with the world. By contrast, current computers and AI systems learn passively, through the "big gulp” of data. While I agree that large language models may already pass certain versions of the Turing Test— though, echoing Prof. Harnad, perhaps it might be “cheating”— Turing’s proposal of child machines that learn and adapt offers a more compelling trajectory toward genuine intelligence. I would extend his suggestion further: if such machines were not only capable of learning but also embedded within a human-like physical body, moving through and interacting with the environment as we do (even though they would lack processes like eating), their learning might be qualitatively transformed. This embodied grounding could, in principle, bring them closer to developing something akin to inherent thought.
ReplyDeleteMaya, good synthesis, but perhaps some (if not all) of T4 may be needed to pass T2/T3 too. Some of the robotic body and brain could turn out to require some biological (and even biochemical) components and capacities to successfully pass T2/T3. Some of the "vegetative" functions might have to align with the "cognitive" ones...
DeleteIn his article, Turing explores the idea that a machine may be able to “be the subject of its own thought” as well as generate new programs for itself. Assuming this to be true, I believe that AI might be the key to solving the hard problem. Indeed, I think that the hard problem is likely unsolvable because the scope of its solution goes beyond the capacities of the human brain, i.e. the brain could not understand it. However, if AI were to somehow attain sentience by “evolving” its programs, it might be able to introspect and figure out how it is able to think and feel. Then, perhaps it could verbalize and vulgarize the causal mechanism of feeling down to a brain-friendly fashion, which could partly resolve the hard problem. This is, however, all hypothetical (maybe even utopian), of course.
ReplyDeleteCendrine, but our real, human, biological brains have been through that evolutionary process, and yet we don't yet have a solution to the Hard Problem. There is reason to believe that our brains will eventually be able to solve the Easy Problem. That is just a matter of reverse-engineering, Turing-Testing and 3D-Printing, after all. But with the Hard Problem that is not so evident -- because only feelers feel, and the only feeling they can feel is their own. So perhaps we can never know how or why (or even whether) a successfully reverse-engineered, Turing-Indistinguishable T4-passer feels. Perhaps only the T4-passer can know for sure.
DeleteBut perhaps we can agree with Turing that T4 Turing-indistinguishability would be close enough...
In section 7, Turing discusses the future prospect of “learning machines” and the “process of trying to imitate an adult human mind,” and he does so largely through the lens of evolution. While Turing seems to more or less concede (in The Argument from Consciousness section) that his test would not satisfactorily address the problem of “solipsism” (or more accurately the problem of other minds), considering the development of “thinking machines” in parallel with evolution has left me more convinced that we will someday develop a conscious/feeling machine, whether we can “prove it” or not. If we assume that sentience arises from some aspect(s) of the physical body (as we all do for the sake of science), then we can assume that the most basic level of sentience emerged at some point over the course of evolution via some physical mutation. Since intelligent design is not random, it seems like it should sort out beneficial mutations much more efficiently than natural selection (assuming we can manipulate the appropriate scale of the physical world). If our understanding of evolution is true, then given indefinite time, resources and motivation, does the production of sentience in a man-made machine not seem almost inevitable?
ReplyDeleteLiam, very good points, but some some of them may require some more reflection:
DeleteSince, because of the "other-minds problem," we (the "intelligent designers") cannot mind-read, neither can Darwinian evolution: Turing Indistinguishability and the Blind Watchmaker
That does not mean, however, that sentience cannot evolve because of some sort of adaptive benefit it confers on survival and reproduction. But the selective guidance that that confers on survival and reproduction is... survival and reproduction! So we "intelligent designers" would have to "evolve it" in our experimental genomes the same way the Blind Watchmaker did -- by starting with a form of life that we assume (on what grounds?)is not sentient, and then either wait for natural mutations to recapitulate how sentience first evolved (which would not only take as long as it did for the Blind Watchmaker, but it would not tell us anything (and would face obstacles similar the ones faced by origin-of-life experiments: the impossibility of generating abiotic synthesis in a biosphere already suffused with biosis).
Or intelligent-designers would instead have to try to do systematic genetic engineering on the non-sentient genomes, and observe which manipulations improve survival and reproduction. But how would we intelligent-designers know that a beneficial effect of our genetic manipulation on survival or reproduction was due to having generated sentience, rather than some other adaptive benefit?
We are as blind as the Blind Watchmaker, because of the Other-Minds Problem. Moreover, it's almost certain that our starting non-sentient life-form would have to be before the vertebrates, who are all almost certainly sentient. In fact, the starting life-form would be low down even among invertebrates, for which evidence already suggests that cephalopods, gastropods, and insects are already sentient: Six-legged suffering.
So, whether the intelligent genetic engineering begins with clams, or jellyfish or sponges, or even lower, with bilaterians, amoeba or other microbes, it's hard to imagine what genetic traits it would be intelligent for us to manipulate, and what would be our "marker" of the fact that any enhanced survival and reproduction caused by our genetic engineering had occurred because our manipulation had produced sentience rather than some other adaptive benefit.
So I'm not sure that intelligent design has a time-advantage over natural selection (which had had billions of years to evolve life, and at least another 500 million to evolve sentience) when the target is an unobservable trait that has enhanced fitness by launching sentience rather than something else.
In Week 4, we will be treating something analogous to this -- mirror neurons and "mind-reading" capacities. In that case, Darwinian evolution could have conferred on some social vertebrates the capacity to "read the minds" of their conspecifics, or even of other species, despite the other-minds problem and the blindness of the Blind-Watchmaker. I won't explain it to you till Week 4, to give you or others time to ponder how that could be possible.
“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?”
ReplyDeleteI like this passage because it shows how Turing takes a really messy and complicated question: “Can machines think?”, and turns it into something you can actually somewhat test. Instead of arguing about definitions, he suggests seeing if a machine can fool someone in a conversation just as easily as a person can. If whatever system we call a "machine" can imitate a human well enough to fool an interrogator, then for the purposes of the test, it counts as "intelligent". It’s clever because it makes the idea of machine intelligence practical instead of just philosophical. However, I think this still leaves open the deeper question of whether the machine truly understands. Personally, I find it wrong (or at least too simple) to assume that being able to imitate human performance means understanding. A parrot can be trained to imitate human speech by saying “hello” whenever someone walks into the room, but does it actually understand what “hello” means?
Jad, please always read the other comments (and my replies) that have been posted before yours, to make sure you are not just repeating what others have already said (e.g., about "imitation."
DeleteTuring’s response to the Lady Lovelace objection (i.e. that machines can “only do what we tell them”) raises questions about free will and originality, especially in the passage: “[…] "There is nothing new under the sun." Who can be certain that "original work" that he has done was not simply the growth of the seed planted in him by teaching […].” This touches on whether human thought is autonomous or just shaped by experience. I don’t want to get stuck in unresolvable philosophical questions (like free will), so here’s my main point: to meaningfully ask “can machines think?”, we first need to lay out the assumptions we make when we say “humans thinks” and give a proper definition of what we consider it to be. That definition has to go beyond “we can feel what it feels to think”, because, due to The Other Minds Problem, it seems that we can never verify if machines feel anything (i.e how can we know if other beings feel?)
ReplyDeleteTuring also mentions machines “never taking us by surprise,” but being surprised doesn’t necessarily imply true novelty, e.g. forgetting something and being reminded can also feel surprising. Maybe I just misunderstood Turing’s refutation of the Lady Lovelace's objection.
Turing’s refutation of Lady Lovelace highlights that what we might consider as ‘original’ or ‘creative’ in human thinking and reasoning may itself be the product of our own prior experiences and things that we have internalized, so the boundary between human and machine creativity is blurrier than it seems and yes I agree it depends on what we would define as “thinking” for humans, because thinking doesn't necessarily mean original. I’d add that this also supports Turing's behavioural approach. Instead of trying to settle unresolvable questions like whether a machine thinks or feels, we can focus on whether it can perform tasks in ways indistinguishable from humans. This allows us to evaluate ‘thinking’ in a way which sidesteps the problem of Other Minds while still giving meaningful criteria for determining “intelligence”.
DeleteEmmanuelle, we know what it feels like to think, but we don't know the causal mechanism that produces the thinking, nor what produces the capacity of thinkers to do what they can do.
DeletePlease read the replies to the other skywritings on novelty, etc. The short answer is that novelty is irrelevant if you understand the power of Turing-Indistinguishability as the basis for Turing-testing.
Lauren, good grasp.
"If we substitute "laws of behaviour which regulate his life" for "laws of conduct by which he regulates his life" in the argument quoted the undistributed middle is no longer insuperable. For we believe that it is not only true that being regulated by laws of behaviour implies being some sort of machine (though not necessarily a discrete-state machine), but that conversely being such a machine implies being regulated by such laws"
ReplyDeleteI feel that I can almost buy that argument. Maybe it is due to a misunderstanding that leads me to not fully be into it, however I don't see the reason to split laws of behavior and laws of conduct other than to try to counter the original argument. At the end of the day, they are both forms of thinking or "feeling" as we put it in class. I don't think taking the similarity between the "laws of behavior" being similar to machines makes for a compelling argument.I think that you could very easily say that the absence of laws of conduct in machines can be taken as a way to counter Turing's argument. Although, I don't believe that humans can't be machines, but I am having trouble piecing what Turing is saying here.
Nevertheless, I believe that a stronger case is made later when he proposes imitating the mind of a child instead of that of an adult. I believe that it might be impossible to simulate an "adult" machine. An adult comes with tons of experiences that might affect how they think and , most of them, might not even be aware of how themselves. As we have discussed in class, we can't really get to know how we come to "feel" through introspection alone. I believe that we know almost as much as about what is going on inside a child's head than those of other adults. An infant at least get us as close to a blank slate as we can get. Thus, if we are able to simulate an infant in a machine and then it eventually evolves into that of adult passing the Turing test, it would be the most compelling argument for a "human" machine.
Jean-Rémy, your comment was almost 300 words (not counting the quotation): Please keep it below 100 words.
DeleteTuring was just talking about causality, whether we do what we do because our brains cause us to do it, or learning from prior experience causes us to do it (traffic fines if we speed), either way it is the underlying causal mechanism we want to discover and test. (He does give a hint that he is not a computationalist: did you detect it?)
He is not talking about feeling, because he has already announced that his method cannot test feeling (unobservable), only doing (observable).
I am very curious about finding the hint that Turing is not a computationalist, so I read the Objection 8 again to look for it. From what I understand, Turing agreed that human behaviour will never be fully predictable because even simple rules of conduct like traffic rules cannot predict what will a human do in all possible situations on the road. Does this mean that, while it can help understand many underlying causal mechanisms of cognition, Turing believes that a machine/computation will never fully capture human cognition, thus hinting that he is not a computationalist?
DeleteAnne-Sophie "...being regulated by laws of behaviour implies being some sort of machine (though not necessarily a discrete-state machine)..."
DeleteI had a suspicion that turing wasn't a computationalist, however, I didn't catch that hint on first read. If i understand correctly, computationalists believe that humans are a discrete state machine? Thus, Turing mentioning that it isn't necessarily one, hints at his belief against computationalism. is that right?
Delete
ReplyDeleteTuring makes a fair point when comparing adult human minds to child-like learning machines, highlighting that “processes that are learnt do not produce a hundred per cent. certainty of results; if they did they could not be unlearnt”.
Unlike digital computers, humans operate through flexible and probabilistic thinking patterns allowing us to quickly adapt and learn (or unlearn) in novel situations. This adaptability often leaves room for error or “fallibility” which is not inherently fixed or programmed but simply human nature. Computers, on the other hand, execute specific algorithms that yield the same output with complete certainty when given identical inputs. Faced with a puzzling scenario that involves complex cognitive functions such as reasoning or problem-solving, a computer would thus not deliberate in the human sense (i.e., "what should I do?") but simply follow its rules (i.e., "what has been programmed for me to do?").
Grace, if a T2 candidate said exactly the same thing whenever you said exactly the same thing, it would would immediately be distinguishable from a real person, so it would fail T2. Does ChatGPT say exactly the same thing whenever you say exactly the same thing?
DeleteEveryone: It is now 2025, 75 years after Turing wrote this paper. The time has passed for saying the same things that were being said 75 years ago about what computers cannot do. Please try it with ChatGPT or Claude or Gemini before saying they can't.
And in testing whether or not computationalism (C=C) is correct, it is not the computer but the T2-passing algorithm that is being tested. (Why?)
It was both challenging and insightful to read Turing’s classical paper because I had always been reticent about the idea that computational machines could recreate human thinking. Similarly to the objections that Turing addressed, a part of me felt scared (and almost offended) that human thinking could be reduced to just rule-following manipulation of symbols. Then, after reading his paper, I understood that Turing acknowledges the complexity of human motivation, feelings, and experiences, but that his goal is to reduce our uncertainty about human cognition with computation (just like the sandwich machine).
ReplyDeleteAs he mentioned in Objection 5, it would be “idiotic” to program a machine that can feel the enjoyment of a delicious dish. His point is to rather understand what are the environmental factors that make humans come up with their reasoning and thinking.
I think Turing came up with a very clever method that avoids the subjective and inefficient homunculus and introspection, and that instead measures scientifically and objectively cognitive processes. The paradigm consists simply of a rule-following machine reading, writing, and storing symbols, but it allows us to recreate and determine the experiences and learning needed for us to think all of the complex thoughts we have in our brain.
Please let me know if my understanding of Turing’s main idea is accurate and if my explanation was kid-sibly.
Anne-Sophie, that was kid-sibly, but at 223 words, too long. (Please don't make it more than 100 words. I have to sleep!)
DeletePlease be sure you read the other commentaries and replies before you post your own commentary. Especially the replies, because they contain a lot of core course material.
What is computation? What is reverse-engineering? What does the Turing Test Measure, and why?
Sorry about the length and repetition. I understand better now what you expect us to do in skywritings.
DeleteTo answer your questions, computation is symbol manipulation based on previously programmed rules, such as “if you read ‘+’, add the number stored in position 4 to the number stored in position 3 and write the result in position 5”. Reverse-engineering is finding underlying causal mechanisms of something, or in other words, finding the necessary input and rules to create a certain output. The idea of the Turing test is to develop a computational machine that manipulates letters and numbers according to rules that will allow it to produce verbal responses indistinguishable from those of a human. So, the Turing Test measures the discrete rules and steps needed for this machine to create human-like responses, i.e., thinking.
I think Turing dismisses some of the arguments against the possibility that machines cannot think too quickly. For example, the Argument from Continuity in the Nervous System is personally quite persuading. Indeed, by the simple fact that the nervous system is not based on a binary system and can also be graded, it seems that it would be impossible for machines to think like we do because their hardware is too different from our brains. Even if the output is the same in both cases, I am skeptical to say that machines think and if they do it would be very different to what we intuitively believe thinking is.
ReplyDeleteIn this paper Turing asks if machines can truly think and comes up with the imitation game as a practical way to test intelligence and compare differences between humans and computers to highlight the differences between them. It was nice to see that he had taken into account the human experience of living - experiencing emotions - but said that machines may be able to replicate this in the future. This is true too to some extent - when you chat with e.g. ChatGPT or Claude, they almost replicate the way you speak to give you a similar response you might expect. Turing responds to the argument for consciousness by pointing out that this leads to solipsism - if you had to feel what a machine feels to know it thinks, you couldn’t ever be sure about other humans either. I think he moves past this a bit quickly. While I get his point, it raises the question - can intelligence really be judged purely by imitation, or is there something deeper about understanding?
ReplyDeleteI wonder if you could test the difference between real thinking and mimicry by pushing a machine with deeper questions. At some point, pure symbol-shuffling should break down unless there’s actual reasoning. Turing hints at this problem—since we can’t access another person’s mind either, we fall back on behaviour as the measure.
ReplyDeleteIf behaviour is enough to show thinking in machines, imitation might be the test. But if thinking means reasoning, then the imitation game is just a starting point. And maybe even mistakes matter—people can be wrong and still be thinking. Could a machine’s errors also be evidence of thought? Or am I thinking about this the wrong way?
From what I understand, one of the most forward-looking parts of Turing’s proposal was his belief that machines could learn rather than just follow fixed programs. He compared this to raising a child, where the machine starts simple and is trained through feedback. This was right because it anticipated modern machine learning, where systems like neural networks improve by adjusting from experience rather than being hand-coded. But the problem is that Turing treated learning as enough to reach human-level cognition. As later debates like Searle’s Chinese Room show, learning rules still doesn’t explain how meaning or understanding arises. So Turing was right to emphasize learning, but wrong to assume that training alone would solve the full problem of cognition.
ReplyDeleteLady Lovelace states that "the Analytical Engine has no pretensions to originate anything. It can do whatever we know how to order it to perform". While reading this, I can understand why she saw machines as purely mechanical, with no creativity, yet Turing's response resonated with me. In fact, he responds that machines can indeed take us by surprise because our own predictions are often incomplete or mistaken. This brings back the issue of 'novelty' and whether it is linked to intelligence. I don't think novelty particularly makes computation more intelligent. However, I think it somewhat makes it more human, which relates to Turing's argument about indistinguishability. Thus, if machines can surprise humans in the same way humans do, this makes machines pass the Turing test, as they convincingly imitate human behavior. Following Turing's logic, machines can indeed think as they can interact in ways humans recognize as intelligent.
ReplyDeleteThis reading talks about the relationship between AI performance and human behaviour. If machines like LLMs can achieve a 73% win rate in the Turing test by simply adopting a convincing human-like persona, doesn't that imply that the test is merely measuring humanlikeness? What if that's what us humans are fundamentally doing too?
ReplyDeleteOur brain is equipped with a mirroring system, using mimicry to help understand other's emotions and actions. If we are biologically driven to mirror, then maybe the LLM is simply a reflection of the human mind, and thus a true Reverse Turing Test. The model would just adopt the peresona I promt it with, thus mirroring my expectations. I suppose the difference between humans and and AI is then simply that my own mind and concepts are grounded in the physical world, whereas tthe machine output is solely based on imitation
In Computing Machinery and Intelligence (1950), Turing introduces a behavioural criterion to replace the question “Can machines think?”: machines can achieve functional parity if its verbal performance is indistinguishable from a human’s. prof Harnad extends this into a methodological challenge — understanding what makes this performance possible requires investigating reverse-engineering.
ReplyDeleteReading Turing alongside Jones & Bergen and Sejnowski highlights how much the benchmark has drifted. Large language models may appear to “pass” T2, but they rely on the “Big Gulp” of human-produced data rather than autonomous grounding. This is a behavioural, not a causal achievement; Sejnowski’s “Reverse Turing Test” explores this and demonstrates how humans confuse fluency for understanding.
We can empirically test this: if comprehension is more than probability, then scaling data can’t substitute grounding. A system that interacts with the world and updates through sensorimotor feedback would test if Turing’s criterion still distinguishes explanation and imitation.
In this reading, Lady Lovelace claims that machines “can only do what we know how to order them to perform.” Turing, however, argues that even within those limitations, machines can still produce surprising or seemingly original results. He says that machines often take him by surprise, during programming, as he does not always make “sufficient calculations to decide what to expect them to do.” However, this is not an adequate response to the variant of Lovelace’s objection he raises. It does nothing to demonstrate that machines can generate genuinely surprising or creative outcomes; rather, it highlights human ignorance regarding the complexity of the systems we design, and it does not bring us any closer to answering the question of whether machines can think or pass the Turing test.
ReplyDeleteTuring argues that making a ‘child machine’ that learns and eventually becomes capable of conversing indistinguishably like an adult human would be the easiest way to achieve human level conversational ability through text. Although LLMs today using the ‘big gulp’ is said to be considered cheating, I argue that it is the simplest way to simulate the world syntactically without needing senses (sight, hearing, etc); a plethora of information, similar to having lived million of lives, that a machine would have to sort through and output to reply in conversational interactions is likely the only way to achieve that without making the machine ‘sensuous’.
ReplyDelete***EVERYBODY PLEASE NOTE: I REDUCED THE MINIMUM NUMBER OF SKYWRITINGS. BUT THE READINGS ARE **ALL** RELEVANT TO AN OVERALL UNDERSTANDING OF THE COURSE. SO, EVEN IF YOU DO NOT DO A SKYWRITING ON ALL OF THEM, AT LEAST FEED EACH READING YOU DO NOT READ TO CHATGPT AND ASK IT FOR A SUMMARY, SO YOU KNOW WHAT THE READING SAID — OTHERWISE YOU WILL NOT HAVE A COMPLETE GRASP OF THE COURSE TO INTEGRATE AND INTERCONNECT FOR THE FINAL EXAM.***
ReplyDelete