Can Machines Think?

Can a machine really think—or just fake it well enough to fool us?

In this episode of American Socrates, we dig into Alan Turing’s famous test for machine intelligence. The idea is simple: if a machine can carry on a conversation indistinguishable from a human, maybe it counts as “thinking.” But is that a good standard—or just clever imitation?

We explore why the Turing Test might be too generous, letting in machines that don’t really understand anything. And too strict—shutting out minds that don’t look like ours. Along the way, we meet Ada Lovelace, who saw these problems coming way back in the 1800s, and John Searle, whose “Chinese Room” challenges the whole idea of AI understanding.

If you've ever wondered whether machines can think—or what “thinking” even means—this one’s for you.

Transcript

If Hollywood is to be believed, what started with Furbies in the late 1990s ends with Terminators in the not-too-distant future. I remain unconvinced that this is truly where things are headed. AI could be scary, and it can do real harm, but is it actually thinking? Is it actually conscious? And if so, does it have desires of its own? I mean, does my laptop get sad when I turn it off or angry when I don't take its suggestions? I don't think so. I doubt that AI has an ego and certainly no emotional response. We already interact with bots more than we think. When you're chatting online, debating movies or something, and you're convinced the person on the other end is sharp, witty, and fully engaged, but there might not be a person at all, just AI. Would that change how you feel about the conversation? And is AI really a machine thought at all? Or is it just a sophisticated way of communicating and collaborating with other human beings? Welcome back to American Socrates. I'm your host, Charles M. Rupert. Today, we're going to be discussing the growing relevance of AI and modern life. Well, we're going to start by taking a look at Alan Turing's significant paper in 1950 called Computing Machinery and Intelligence, where he laid out the first ideas of how to determine if a machine is capable of thinking. What we want to know is, how does understanding machine intelligence shape our understanding of human intelligence? Can we learn more about the way we think by looking at the way machines supposedly think? Alan Turing proposed a simple test for understanding whether a machine was thinking or not. It's now referred to as the Turing test. What he did was turn the spotlight on the nature of intelligence itself by creating a sophisticated analog. It's not so much can the machine think, but could a machine ever convince you that it was human? And it if it could, does that mean it's capable of thinking on its own? So in this episode, we're going to really take a look at the Turing test and explore its strengths and its flaws and tackle the deeper questions about what it means to think, both for humans and for machines. Let's see if we can unravel the mystery behind artificial intelligence. The Turing test is actually based on a 19th-century parlor game known as the Imitation Game. It's the sort of thing that the Victorians would do when they had an entire afternoon to kill. The imitation game works with three players. The first player is a male, and the second player is a female, and the third player can be a member of either gender. The male and the female, we'll call them player A and B, go into another room. From this point on, they cannot interact with the third player, player C. The only way that they can communicate is by slipping notes underneath the door. The same player will write answers for both of them, so there's no way to tell whose handwriting it is. The goal is for player C to ask questions of the players A and B. They could ask either one, whatever they want, and as many questions as they need until they finally guess what the gender of each one is, as player A, the man, or is player A, the woman? And so this sort of like gender bending game was meant as a test to figure out like, which person you were talking to. Now, Trin thought that this could be used as an analog for machine intelligence. Can a machine play the imitation game on as well as a human so that the interrogator would not know whether or not they were talking to a human being or a machine? The way Turing describes it, we would replace one of the two players in the other room with a machine. And some of the answers would be with to the human and some of the answers would be to this computer. If the interrogator cannot tell that they're talking to a computer during this game and wins the game just as often as the human would win the game, then there's no way that must be evidence that the machine is thinking. There's no way that a machine could convince us of its, you know, gender or whatever if it wasn't capable of some kind of thought.. The idea behind this test is that we don't actually know what thinking is. I mean, it's really hard to describe. What is a thought, what is thinking? And so Turing's observation is that what we do when we determine a person to be thinking is that we base it on their interactions with us. You know, I never observed another person's thoughts, and neither do you. You have your thoughts, but you're kind of alone in your head. What you do is you see that they speak, that they move and act in certain ways, that you determine to be purposeful. And because of that, you determine them to be thinking. They are something that thinks. So Turing believes that what was gender in the imitation game could work for thinking in this slightly modified version. Turing claims that if the machine can successfully imitate a human being to the point where it's just indistinguishable, then can it not be said to be thinking? So what Turing has done here is give us a functional definition of intelligence, based on performance, not some sort of sense of inner experience. The machine doesn't know, and we would never know if the machine was doing its own internal thinking. But just like when we see other people and the performance that they give us, the words that come out of their mouths, the behavior that they give us, and we can make a judgment call on that, that they are, in fact, intelligent and that they are, in fact thinking, we can do the same thing with a computer. If it can mimic what humans do to that point, then it must be doing some kind of thinking. Now Turin believed that within 50 years of that paper, that is, by the year 2000, we would have invented computers with sufficient strength to be able to mimic humans in such a such a point that he believed, I think it was 70-30 that they would win 70% of the time. And that he believed was a sufficient criterion to say that they had passed the Turing tests, that humans were successfully concluding that they were thinking 70% of the time. But by the year 2000, no computer had passed the Turing test. We still... don't have computers that completely pass it. We have all of these artificial intelligence agencies, you know, from Apple to Amazon to Google. And all of them still give some wonky answers every now and then. It's pretty obvious after talking to one of them for a few minutes that you are, in fact, talking to a machine and not a human being. So it's possible there might be something wrong with the terrine test. And that's what I would like to turn to now. I would like to try and take an analysisysis of the turning test to see if there isn't some sort of inherent flaw in the test itself. One of the first objections to this actually comes long before Turing is from the daughter of the celebrated poet Lord Byron, the Lady Lovelace. She had a correspondence with Charles Babbage, the inventor of the analytical engine, the first true attempt at building a computer in the modern sense of something that was like programmable and could work in many different capacities and didn't have like a limited sort of capacity for doing certain kinds of calculations. He never really got it to work, not in the sense that we think of a computer working today, but the idea of the analytical engine is what intrigued Lady Lovelace. What she proposed was that the machine could never really be thinking because nothing that ever came out of it was more than what was put into it. Now, this idea is a little tricky to understand. What she's basically saying is, is whatever we program into a computer is all that we ever get out of it. And so if there's any thinking that's going on, it's really somehow us doing it. You know, the machine cannot generate new thoughts. There's nothing novel about it. Now, Turin responds to this idea in his paper, where he sort of twists Lady Lovelace's objection to say that the computer is itself not surprising, rather than what I believe she had in mind, which was that the computer doesn't generate novel thoughts, he suggests that the reality is, is that we just don't find anything surprising about what computers do because we programmed them. He contradicts it by simply saying that he's been surprised many times by what computers do. You know, when the programming goes slightly awry or it's doing something he wasn't expecting that response, he finds to be surprising. And so he discounts her as being wrong. But if we get back to her original complaint, the idea that it's not about being surprising. It's about that there is nothing that a computer can know that we didn't know first. It's not finding out any new information on its own. It's waiting for us to find it out and then maybe surprising us in and of the sense that it's getting our information from somewhere else or from someone else, but always from a human, always from somebody. This seems to suggest that computers do not in fact think in and of the sense that they can't seem to generate ideas that they haven't been explicitly programmed for. You know, with AI, we might say that it seems like they're thinking because of the great volume of information they can respond to us with. But the reality is, is these large language models are simply trolling the internet for massive amounts of information and it didn't create any of that information. Human beings created all of it. It's simply looking over all of the possible languages that human beings have written and piecing together something to respond to us with. Sometimes it gets it right. Sometimes it gets it wrong. But that's more the algorithm of what it's selecting and why it's selecting that. The reality is, it's not creating its own thoughts. It's not doing what even a baby can do or even an animal can do. It's not having an opinion of its own. That's Lady Lovelace's objection here. So the first flaw in the Turing test could be that he misunderstood or at the worst, misrepresented Lady Lovelace's objection, and that her problem with analytical engines is that is more compelling than he gives credit to. A second problem worth noting here comes from the American philosopher John Searle. In Searle's Chinese Room argument, he points out that the machine might simulate understanding without actually being capable of comprehension. The thought experiment goes something like this. Imagine a person inside a locked room, similar to the imitation game that we were discussing earlier. In this room is a series of rule books. All of the rule books have certain codes that if they receive certain inputs, you look up whatever that input is, and then it tells you what output to give. Now, what the inputs and outputs all are are Chinese characters for Mandarin Chinese. The person inside the room doesn't necessarily know Mandarin Chinese. However, notes come in under the door from someone outside the room who speaks Mandarin Chinese. This person then picks these notes up and looks at what's written and goes and uses the rule books to figure out which parts they're supposed to respond to and what they are supposed to respond with. They then write out the characters that they are supposed to respond with and slide the note back under the door. If this is compelling enough, the person outside the door would be convinced that the person inside the room is or does speak Mandarin Chinese. They're giving all the right answers to all of their questions. But do they really understand Chinese? They are giving the correct answers, but they're only giving them because of whatever the rule book said. It seems more likely that the rule books seem to understand Mandarin Chinese and were probably written by someone who spoke Chinese. This person is just going through the motions of executing what someone else had already written some time before that. So Searle's point here seems to be that we can have an imitation game-like scenario in which the person is convinced that this person is doing the thinking or, in this case, understanding the language when, in fact, they don't. They do not understand the language. It only appears that they do. Now, this idea sort of turns on one of the problems with Turing's test to begin with, and that is the idea that it is based on an analogy. The computational model of computing intelligence that Turing is sort of using here it was like later developed by like Hilary Putnam and followers of his students. That model is very functionalist in its design. We have a very functionalist idea of what it means to be thinking. Very briefly, functionalism just means we're going to define things by their function. All chairs are chairs if you can sit on them. It doesn't matter what shape they are, what materials they're made out of, the look of the chair, the feel of the chair. None of that matters. All that matters is, can I sit on it? In that case, a cut down tree stump and a throne are essentially both chairs. If we look at the mind then and thinking purely in functional terms, we have a certain description of it. It performs these sorts of functions. It performs these sorts of behaviors. And when we see those types of behaviors, that must be evidence of thinking, because that's all it is. So Turing is using this, where he's equating thinking to these functional outputs, this sort of behavior. Does it speak right? Does it act right? Does it say the right sorts of things? But just because something behaves like a thinker would doesn't mean it's necessarily thinking. And that is what Cyril is getting at here. The idea simply being that it could be really just imitating thinking. It's faking it in all the right ways. It's appearing to be thinking, but not thinking in reality. For guys like Turing, there's no way to tell the difference because functional definitions as functional definitions will not allow you to distinguish between something that appears to be functioning and something that is what it is. So one of the ways that we can try to understand this analogy is by looking at what is being equated between the two sides in this debate. There's two different versions of it. We could say. There's a weak equivalence or there's a strong equivalence. And we need to know which one we're really dealing with here if we're going to analyze machine thinking and the turnuring test properly. An example of strong equivalence would be to say that 5 plus 7 and 9 plus 3 and that 4 times 3 are all completely equal. They're all the same number. They're all 12. And so because they're all all 12, they're all the same number. Those are the same thing. They're just different ways of expressing it, but those all mean exactly to the same thing. Even as a function, they're all the same number. Weak equivalents would be more like the example I gave earlier about the chairs, a tree stump, a camp chair, and a leather sofa, were would all be chairs in and of the sense that you could sit on them, even though they look completely different. They feel completely different. They're made out of completely different materials. But from this one very particular point of view, they do have a similar function. And so therefore, they could be the same thing or thought of as the same thing. That's just a weak version of it equivalence compared to, you know, five plus seven and 9 plus three and so on. So our question now becomes which kind of equivalents are we really dealing with here, comparing human brains with silicon chips? Is it a weak equivalence or is it a strong equivalence? If it's a weak equivalence, we would say that the machines and the human brains may achieve the same results, but they're arriving at those results through vastly different methods. It's not the same process that gets us the same behavior. Yes, in the end, the behavior definitely looks the same, but it's arrived at in such a different way that we would not say that the computer is necessarily thinking. If it's a strong equivalent, on the other hand, the machines in the human brain are not only achieving the same results, but they're doing so in a very similar manner, which means they're doing it both by thinking. So the objection to the Turing test here is that it can only really prove weak equivalence. That's what Cyril's trying to say. There's only this weak equivalence that we can say the Turing test gives us. We don't know that the computers are in fact thinking. We would just have to say that they are capable of mimicking thinking. The problem with this should be obvious to us. You know, a voice recording sounds like someone thinking, but does that mean that a tape recorder or an answering machine or, you know, a voicemail or something like that is actually thinking? Probably not. You know, there was some thinking going on there, and that was the thinking of the person who recorded the message. So there always seems to be this here human behind the electric curtain somewhere. That's what Lady Lovelace was getting at. And that's also what John Serril is getting at. This idea that there always seems to be a human behind the machine machine and that the machine is not generating anything on its own. It's not doing anything on its own. It's worth noting, though, in the turnuring test defense, that despite the best efforts to create a series of prechosen responses to convince people that AI is in fact human, we never really have prodigiously succeeded. There's always been some doubt here. So it might actually prove the test by simply saying that computers don't think. No computer is capable of thinking because they all fail to generate anything novel or anything new. And on some level, that's going to make sense to us. You know, a computer shouldn't actually think like a human being. It doesn't make sense that it would. Computers do things differently. If a computer was thinking, if my laptop, for example, was thinking, I would imagine that it would have no problem answering certain questions that would stymie me. For example, if you were to ask me, what is the square root of 144,632, I I would have no idea. And it would take me quite a few minutes to try and figure out what that is. Whereas if you type that into the computer, it would spit out an answer in a matter of seconds. Now, imagine we reversed the Turing test and had computers testing humans for intelligence, and it started off with a simple question like that, something that any computer ought to be able to calculate in a matter of seconds and watch human beings thumb around and go, uh, it would seem to the computers then that human beings are not intelligent. So it's possible that we're having some kind of, you know, false negative here, that computers are, in fact thinking, even though we can't seem to derive it from the term Turing test because they're not answering the kinds of questions we would want to answer. Another way to think about this is, you know, let's imagine some human-comparable-intelligent dolphins. You know, would their thinking ever really render them capable of convincing a human that they were human? They wouldn't think like a human. They wouldn't act like a human. And exactly the same problem in reverse. Could you ever convince a bunch of dolphins that you were a dolphin? Probably not. They would have experiences that are different from you. They would be able to talk about things like echolocation, the experience of echolocation, which you were wouldn't know how to. I don't expect my laptop to be able to tell me what chicken tastes like because it doesn't have a mouth and it doesn't eat chicken. But I do expect it to be able to answer math problems rather quickly. So perhaps the real problem with the Turing test is that it makes intelligence dependent on being human and sounding like a human. That's probably some sort of anthropocentric fallacy. There's all sorts of different intelligences out there, an animal intelligence, a machine intelligence. And they wouldn't look anything like human intelligence. If that's the case, the Turing test fails in both directions. It's susceptible to false positives in and of the sense that it might take tell us a machine is thinking when in fact, it's only appearing to think as Lady Lovelace and John Cyril have pointed out, or it might give us false negatives in which the machine could very well be thinking, but the Turing test wouldn't convince us that it's thinking because it doesn't sound human enough. So in the end, I think that you have to think of artificial intelligence, at least as it is currently at, as less like human thinking. There's no comprehension, there's no deep thought there. It's really a sophisticated organizing program for taking large amounts of data and repackaging it very, very quickly in such a way that we can't see what it's doing. The complexity of this system then gives us the illusion that there is a thought or a person or some sort of intelligence behind it all. It's very similar into the way Oija board works, where everyone puts their hands on the indicator and moves the Ouija board around. But since the movements are very slight and almost imperceptible, we don't realize that it's us that's moving it. Not any one person who has their hands on the Ouija board feels themselves moving it, and yet it seems to move. And so there's this seemingly spiritual element that is moving the Wija board around when actually the only intelligent thing that was there were the people who have their hands on it. It was us the whole time. AI is very much like that. It's like a giant Ouija board. It's really us talking to ourselves through this very sophisticated method that gives us the impression that there's nobody else there, that we're not talking to anybody else. When, in fact, we really are. I think that's actually one of the more beautiful parts about AI is that it's a very sophisticated and fancy way of collaborating with other human beings. The problem with it is, is you don't often know who or what you're collaborating with, and there's no way to acknowledge or understand who you're collaborating with. Because it's taking its information from something like the internet, it's possible that you are dealing with all sorts of human biases, with bad actors, with disinformation, and many other factors that make the experience less than pleasant. If we could somehow fix that, it might be one of the greatest gifts to humanity, this ability to interact with each other, and to cooperate in a technological universe that makes the collaboration with millions, if not billions of other people possible, all in a single moment, encourage you as you begin interacting increasingly with AII, to take a few moments and think about how this machine works and what it's capable of and what it's mimicking and the millions of people behind it. If you start thinking about this as a way of interacting with other human beings, you'll have more appreciation for your species, as well as some deep-seated fears about us as well. If there's anything really scary about AI, it's us. It's human beings. It's not some sort of alien logic and some sort of machine thinking. We don't have that yet. And I'm not sure that we ever will or even want to try and work on that. Nobody is trying to develop intelligent systems that think outside of human thinking. That, you know, doesn't do anything for human beings and are only interested in doing whatever they want to do. That could be dangerous. And those could end up, you know, turning into Terminators or the Matrix or something like that. Thanks for tuning in to American Socrates. Today's episode of Philosophy got you thinking in new ways. Make sure to subscribe so you'll never miss an episode. New full episodes drop every Wednesday. If you enjoyed the show, leave a review. It helps others find us, and it means a lot. And if you know someone who could use a little more practical wisdom in their life, share this episode with them. Want more? Viscrates. buzzrout.com for show notes, resources, and exclusive content. You can also follow me on Facebook, Blue Sky, or TikTok to keep the conversation going. Until next time, keep questioning everything. Join us next week, when we will explore the practical question of learning learning by examining the question, "Who needs grades anyway?"