Kim Mills: In the past few months, artificial intelligence has been all over the news. We’re all trying to figure out how to live in a world where everything from college essays to wedding vows to mental health advice can be outsourced to a chatbot. Meanwhile, the rollout of some AI systems hasn’t been entirely smooth. The National Eating Disorders Association, for example, replaced its human staffed helpline with an AI chatbot, but took it offline after it started giving problematic advice. So today we’re going to be talking about how artificial intelligence works, how AI connects to and differs from human cognition, and how studying human cognition could help researchers improve AI systems.

So what’s going on inside the black box of AI? In what ways are the AI systems we’re using similar to human cognition and in what ways are they different? How is AI changing the way scientific research, including psychological research, is being done? And more broadly, how will it change our lives in the future?

Welcome to Speaking of Psychology, the flagship podcast of the American Psychological Association that examines the links between psychological science and everyday life. I’m Kim Mills.

My guest today is Dr. Tom Griffiths, a professor of psychology and computer science and director of the Computational Cognitive Science Lab at Princeton University. He’s interested in developing mathematical models of human thinking and decision making and in exploring how human cognition connects to ideas from artificial intelligence, machine learning, and statistics. He’s the author of many scientific papers and of a book called Algorithms to Live By, about the parallels between the everyday problems that arise in human lives and the problems faced by computers, and what we can learn from computer scientists about how to solve those problems.?

Dr. Griffiths, thank you for joining me today.

Tom Griffiths, PhD: Thanks, Kim. Great to be here.

Mills: As I mentioned in the intro, AI is everywhere in the news right now, and it seems like it’s becoming smarter all the time. So let me start with a broad question: What are humans still better at than artificial intelligence? And on the flip side, what does AI do better than we do?

Griffiths: Well, I think one way of thinking about it is to think about what the differences are between the kinds of computational problems that human minds have to solve and the kinds of computational problems that AI has to solve. So if you think about humans, we are biological organisms that have limited lifetimes, that have a limited amount of computation, what we carry around inside our heads, and that have limited capacity for communication—we have to do what we’re doing right, now moving our lips and making honking noises in order to get bits from my brain into your brain. And an AI system doesn’t have those same constraints.

AI systems can learn from huge amounts of data. Many of the AI systems we’re dealing with right now are trained on the equivalent of many human lifetimes of data. They can have more and more compute as they need it. You can build bigger and bigger computers. You can wire those computers together. You can have parallel computing going on that’s driving up the amount of computation that those systems can access. And they can just copy the things that one system has learned directly into another system, so they can transfer that information directly.

And so those differences actually set up some of the things that humans are good at and some of the things that AI is good at. So our limited lifetimes mean that we have to be able to learn from small amounts of data, something that AI systems are not great at right now. Our limited computational resources means that we have to be good at using the brains that we have as efficiently as possible, and that’s something that’s still a challenge for AI. And our limited communication capacity means that we need to come up with ways to work together to be able to pool the resources that we have, the data or the computation, to do the kinds of things that no human being can do on their own—creating things like language and teaching each other and culture and societies, all of which are still very human kinds of systems.

Mills: How similar is the way that a system like ChatGPT works to the way that the human brain works? And how is it different?

Griffiths: I think there’s two ways of thinking about that question. One is, when we look at this system as it’s been trained and the behavior that it’s producing, how does that behavior compare to what humans do? The other question is, how does the system learn? How does it get to the point that it’s got to from all of the data that it’s using? And so the second question is a little bit easier to answer because these systems require much more data than a human being to get to the point that a human being gets to by the time they’re say five years old or six years old. So humans in a few years of being exposed to the language around them achieve a pretty high level of competence in using language. And these kinds of systems, it takes much more data, as I was saying before, the equivalent of many human lifetimes of data to get to that point. So if we look at some of these kinds of systems like ChatGPT, you can do the numbers in a few different ways, but they’re trained on at least thousands of human lifetimes worth of data.

And so for learning, it’s clear that the story about learning has to be something different from the way that humans are learning language. The question about how it is that these systems work and whether they’re doing things that are similar to the things that are going on inside people’s heads, that’s still a question that I think people are trying to work out. And so there’s lots of papers that are coming out where they’re doing things like showing that the representations inside these systems can predict things that are going on inside people’s brains. I have a paper like that with a bunch of other co-authors, but there’s a few papers along those lines. And then there’s lots of papers that are probing the behavior of these systems and looking at whether they do the same kinds of things that humans do when they’re given the same kind of stimuli.

And so that’s interestingly a really good project for psychologists. So we as psychologists are used to investigating mysterious systems based just on their behavior. Those systems are normally human beings, but we’re getting the chance to use those same research skills to try and make sense of how AI systems work.

Mills: So it sounds like we’re building the plane while we’re flying it. And I’m just wondering, for example, how it is that some of these chatbots decide, decide in air quotes, to make up answers when they don’t know the answer, why don’t they just come back and say, “I can’t find what you’re looking for?”

Griffiths: That’s a really interesting question, and it gets to the heart of what it means to be a chatbot. So the way that these systems are created, they’re solving a very specific problem, which is how to predict the next word that somebody is going to produce. That’s a problem that you as a human being can—

Mills: Do.

Griffiths: So humans are able to do that. We do that drawing on our knowledge of language and our knowledge of the world and the kind of topics that we’re talking about. And so that’s the same problem that the AI system is trying to solve. So it’s kind of trying to build up, just based on the experience that it has from language, a model of the world that’s good enough to help it to make that prediction of what that next word is going to be.

And so the problem of predicting the next word is not one that sets you up to think about whether that next thing is actually true or not. So now, if you ask this system, “Okay, here’s some text, what’s a likely response to this text?” It’s going to generate something which is a high probability response based on all of the knowledge that it has about language, but it’s not going to be saying, “Okay, is this true or not?” Whereas a human being that you’re interacting with and you’re having a conversation with might be thinking about whether things are true or false.

Mills: So the chatbot, unlike human beings, is not willfully lying, it’s just making a selection based on what’s available to it?

Griffiths: Yeah, it’s not even thinking about truth or falsehood, it’s just thinking about what’s a plausible answer to this question.

Mills: Now, in a recent story in APA’s magazine, the Monitor, you said, “A lot of what’s driving progress is the capacities that these AI systems have, and that’s outstripping how well we understand how they work.” That’s directly quoted from you. What did you mean by that? In what ways do we not understand how AI systems are working?

Griffiths: Yeah, so I made this comment in the context of a proposal that was going around about taking a pause from developing AI systems. And I think that that’s something which is probably impractical given all of the pressures that there are around building these systems. But something that we can do is call for more research looking into how these systems work, how they are similar to people, how they’re different from people. And one of the things that we’re interested in is something that we call alignment, which is how aligned are the ways that these systems think about and represent the world when compared to the thoughts and representations of human beings? Because if you’re used to interacting with another human being, you kind of have a model in your head of how that human being works and what kinds of things that human being is likely to know and what kinds of things that human being is unlikely to know.

And so I think what we need to do is more research where we can work out what are the things that these systems are good at and accurate at, and what are the things where it seems like they’re good and accurate, but maybe they’re not as accurate under the surface, and help people to have an understanding of where it’s probably okay to have this interaction and it’s probably a reasonably safe interaction to be having in terms of the facts that you’re getting and cases where it may not be as accurate in the answers that it’s giving you.

So I can give you a concrete example from our own research where we went in and we asked GPT to give us estimates of how similar different things are to one another. So how similar are two colors, how similar are to musical notes, how similar are two tastes? And what we discovered was that GPT actually has a really good, accurate representation of things like colors and musical notes, that the representation, it’s very aligned with the way that humans think about those things. But its representation of tastes is quite different from the representation that humans have. And so that would tell you, if you were going to ask GPT for advice, it might be good to ask it for a wardrobe advice about what colors are going to go well together, but it might be less good to ask it for advice about a recipe that you’re making while you’re asking about what ingredients would taste well and good together.

Mills: What should the role of psychology and psychological science be in developing artificial intelligence? What is it now and what should it be going forward?

Griffiths: I think there are two ways that we can think about this relationship between psychology and AI, and if you think about it as, there’s the two directions that information can flow from psychology to AI and from AI to psychology. So going from psychology to AI, one of the things that we can do is actually use the toolkit that we’ve developed for making sense of human behavior to really understand what it is these systems are doing, what it is that they know, how well that corresponds to the kinds of things that humans do.

And so psychologists are used to taking systems that are doing all sorts of complex things and just using their behavior, designing experiments that we can run in order to reveal the things that might be going on for a human being inside their head and maybe for a computer inside their circuits to tell us about how it is that that system works and revealing these things that I think are a challenge for AI because a lot of AI researchers have been used to living in a world where you completely understand what your system is doing and are now suddenly in a world where we built these things that are producing complex behaviors in ways that we don’t completely understand and can’t explain just from the underlying mathematics of the system.

Going in the other direction, from AI to psychology, I think we get all sorts of interesting hypotheses about how human minds might work and things that we can explore about human behavior that are inspired by the models that are being used in AI. We also get new tools that we can use for doing our science differently. So ways that we can use those AI systems to design new experiments or to build better models than the kinds of models or theories that psychologists could come up with on their own.

Mills: Let’s talk for a minute about your research. Your work is about making mathematical models of human cognition. What does that mean in lay terms? How do you describe human thinking in mathematical terms?

Griffiths: So where we start in my research is really with this question of, what computational problems human minds have to solve? And that means thinking about when you get up in the morning, you’ve got a series of decisions you have to make about your day, that’s a kind of computational problem, this sort of decision making. You’ve got different options. You have to choose between them. You have to develop a plan. You have to plan out, “Okay, I have to do this before I do this. I have to do this before I do this.” That’s a problem that computer scientists call planning, to you arrive at work. There’s a new concept you have to learn. Someone’s going to give you some examples of things that are examples of that concept. You have to figure out what it is from those examples. You’re going to go and choose where you’re going to eat your lunch. There’s a bunch of restaurants to choose from. You’ve been to some of them before, some of them are new. How do you make the decision about whether you go to something you’ve been to before versus something new?

These are all computational problems. And so what we do is take those computational problems and then think about what is actually a good way of solving those problems. What does the math tell us is the ideal way of solving those problems, and then we use that math to go back and then try and make sense of human behavior. So when we look at things like planning, are the kinds of planning strategies that people use things that we can explain in terms of the mathematics of solving a planning problem. Are the ways in which people learn concepts consistent with the mathematics that comes from statistics about how you should use data to update your beliefs? Are the way that people choose their lunch consistent with something that comes from the reinforcement learning literature in computer science about how to solve what’s called the explore-exploit trade-off. And so in each of those cases, there’s a set of mathematical principles that we can draw on that we can use to go back and understand exactly what’s going on in human behavior.

Mills: You mentioned earlier how artificial intelligence is able to amass all of the knowledge of thousands of generations of human beings. How is AI changing the way that scientific research is done given the amount of data that it’s able to synthesize? How does it change the kinds of questions that researchers might ask?

Griffiths: So as the amount of data that we have, in particular in psychology, the amount of data that we have about human behavior increases, it’s possible for that to get to a point where it goes beyond what it’s easy for a human being to think about. So that creates an opportunity for us to use AI systems to augment the intelligence of humans as they’re dealing with these kinds of data to come up with new ways of making sense of those data to develop new meaningful scientific theories. And so we’ve begun to explore in my lab how we can use those kinds of methods in the context of psychological research.

So we’ve done this in a couple of different ways. So one way is to take a really big data set and take a psychological theory and say, “How much are we missing? How good is our theory really in terms of explaining these data?” And so what we can do then is take a machine learning model, fit that machine learning model to the data, that is, make the machine learning model make the best predictions that it can from the behavioral data.

And then we can compare how good the predictions of the machine learning model are to the predictions of a theory that’s developed by human scientists. And that gives us a kind of gap. It tells us how big is the gap, how much space is there to fill in terms of trying to understand what’s going on? And we can actually use the output of the machine learning model to go back and critique our psychological theory and come up with a better theory. So we call this scientific regret minimization because the idea is that you should care about the errors that you’re making that you could have avoided. And the machine learning model tells you which error you could have avoided by saying this particular piece of data was something that you could have predicted, but you failed to predict. And so that’s one strategy.

The other thing that we can do is actually use the machine learning model to help us search the space of theories. And so we can define a set of theories where we say, “Here’s some sort of psychological principles that are going to go into this, and then how do we combine these together in order to make the best predictions about human behavior?’ And the machine learning model can go off and search through the space of these theories and come up with the best theory in terms of seeing what explains the data. And by doing that, it gives us a tool that we can use to then work with much bigger data sets than we could previously have worked with.

So in one paper recently, we used this strategy to look at human decision making where previously papers on decision making had been used on the order of about a few hundred decisions, choices between different gambles, let’s say, that people might be entertaining. Do you want a 10% chance of this prize or a 90% chance of this prize versus a 20% chance of this prize and a 80% chance of this prize? These are the way that we investigate human decision making. So previous research had used a few hundred of these scenarios to try and understand people’s decisions, but we used more like 10,000 of them. And then we’re able to use a machine learning model to go in and work out, oh, the basic psychological principles we’re using to explain those data are essentially correct, but they don’t apply uniformly across the space of problems. So for some problems, people behave one way, for other problems, people behave another way, and that gives us another kind of thing that we now want to be able to explain about human behavior.

Mills: So in essence, it’s making the science more accurate, is that what you’re saying?

Griffiths: It’s doing two things, I think. One is it’s making the predictions that we’re able to make about human behavior more accurate. But the other is it’s allowing us to explore theories that might be more complicated than the theories we were able to explore before. And that’s a natural consequence of just having more data. So we used to live in a world where we didn’t have very much data about human behavior and so the only theories we could come up with are very simple theories. And as you get more data, you have the opportunity to then explore more complex theories. And these kinds of tools from AI give us the mechanisms that we need in order to be able to explore those kinds of theories.

Mills: As I mentioned in the introduction, you wrote a book several years ago called Algorithms to Live By, about what humans can learn from computer algorithms. Can you talk about that a little bit? I mean, how can thinking a computer help us solve everyday problems?

Griffiths: Right. So my research is about these two perhaps apparently mutually conflicting ideas. So one is what humans can learn from computers, and then the other is maybe what computers can learn from humans. And so the point of Algorithms to Live By was to say that many of the problems that we actually face in everyday life are things that fall into these kinds of categories where in fact, for that computational problem, we know what the solution is and you can just execute that solution. So an example that I mentioned earlier is if you’re trying to decide where to go to for lunch, there’s some places that you’ve been to before, there’s some new place that you could go to. You have a decision you have to make between, do I go to the old place? Do I go to the new place? This problem has been studied extensively by computer scientists. It’s called the explore-exploit trade-off, do you explore the new thing or exploit the knowledge that you have to go to the old thing?

And there are some simple algorithms that you can follow in order to solve that problem optimally. One of these algorithms you can characterize in terms of think about the best plausible experience that you could have at the new place. So be optimistic about the quality of the new place. And if that optimistic estimate is better than your estimate of the quality of the meal you’ll get elsewhere, then you should try the new place. But if you’re pretty confident you’re going to get a great meal elsewhere, then you don’t even need to do that. So that algorithm is now something that you can use when you next have to choose a restaurant. And there’s a bunch of things like this where in fact, just knowing that there’s an optimal solution means that you don’t even have to think about it, and you don’t have to worry about what you do because you know that you’re doing the best thing, even if it doesn’t work out.

And so for people who spend a lot of time thinking about decisions, there’s a little bit of relief in knowing that in fact, you can follow these optimal strategies. That interesting thing though is that there are also cases that are a little bit more complex than these cases where we sort of now have a slightly more challenging problem. For example, say if the quality of the restaurants changes over time, this is a problem which the computer scientists haven’t solved, but that humans are still pretty good at solving. And so the other good bit of news is that your human intuitions are actually pretty good for solving these problems anyway. And so even if you’re not following the optimal strategy, you’re probably doing something that’s pretty sensible. And by looking at those kinds of sensible things that humans do, we can actually work out better algorithms for making computers solve these kinds of problems.

Mills: So it’s kind of being aware of how you’re making the decisions, basically. I mean, our minds are working that way, although sometimes I think you get frustrated and you just flip a coin and say, “It’s Mexican today, it’s not going to be French.”

Griffiths: So it turns out flipping a coin is actually an optimal strategy under some circumstances. So that’s a good way of solving the explore-exploit trade-off, where you, some of the time, are willing to flip a coin. The only thing that I would say is that the rate at which you flip a coin should be decreasing over time. So the longer you stay in a city, you should be taking that random action with lower probability, and then that turns out to be an optimal algorithm too.

Mills: Well, let’s talk about one really big picture question I think a lot of people are asking, especially after that scientist at Google thought that his program was becoming sentient. I mean, will sentience be possible? In your view, do you think that we’re going to be able to make an artificial intelligence that has self awareness?, I mean, what is sentience?

Griffiths: Yeah, that’s a big question. It’s also one that I would say I don’t think about a lot because I’m not sure what difference it makes to the way that we approach these systems or our interactions with them or the way that we think about things. So the first thing I’d say is that we shouldn’t expect AI systems to be like people. And the reason why we shouldn’t expect AI systems to be like people is the constraints that characterize human cognition are different from the constraints that characterize the circumstances under which AI is being engineered. So humans are this weird special case of intelligence that comes from our biological heritage of having limited lifespans, limited amounts of computation and limited communication. And so when you take those constraints away, you’re going to get systems that can intelligently solve problems, they’re not necessarily going to solve problems in ways that look human-like they’re going to solve problems in ways that are different because they’re not operating under the same constraints.

So when AlphaGo was created, which was this AI system which could beat the world’s best human players in the game of Go, that system was able to plan out into the future much further than any human being can plan, it could plan out many, many steps into the future. And as a consequence, what it was doing was it was always taking the move that was guaranteed to give it a small advantage all the way out in the future. And that doesn’t look a lot like human gameplay because humans can’t plan that far out into the future. Humans can only plan over a limited or horizon, and so we do things like setting goals and sub goals and thinking hierarchically and breaking down problems because of that constraint in our cognition. And if you remove that constraint, then you can just solve the problem by planning more.

And that’s kind of the world that the AI systems live in. They’re able to find different kinds of solutions to the problems that we are trying to solve because they’re not operating under the same constraints. And so a lot of the things that we view as characteristic of intelligence are really characteristic of the kinds of computational problems that human minds have to solve. And the AI systems that we’re going to interact with are just going to be different kinds of intelligences. They’re going to be alien in various kinds of ways, and we have to just build our own new internal cognitive models of what these systems are like and what it’s like to interact with them without expecting them to be like human beings.

Mills: Knowing what you know about AI at this point, how do you think it’s going to change our lives in the next few years and in the long term?

Griffiths: I mean, there’s a lot of doom and gloom out there at the moment about AI taking jobs and maybe destroying the world. And there’s a whole spectrum of opinions that you can hear, as well as I think very serious and legitimate concerns about who is being represented in AI and who is represented in the data sets that it’s trained on, and what the consequences are that that has for power in our societies. And so I think all of those concerns are things that are worth thinking about as we interact with these kinds of systems and we see them deployed in our societies, we should be asking questions about what the potential consequences are, who are the people who are involved in this and being represented by it.

At the same time, I think it’s good for us to be optimistic about what the positive consequences of these kinds of systems might be. So for example, in some work that I did with collaborators who have been analyzing the game of Go since AlphaGo was introduced, in that work, what we actually found was that the creativity of the human Go players, as measured by the novelty of the moves that they were producing, was actually increased by the presence of this AI system in their environment. So it’s not the case that any human Go player is able to beat the AI system reliably as a consequence of those innovations, but having this other kind of intelligence enter their community led the humans to explore interesting new directions that they hadn’t explored before and to make their own new discoveries.

And so that’s my hope, is that we’re able to find a way to interact with these systems and to live with them and to integrate them into our communities that allows us to achieve more than we’re otherwise able to achieve as human beings. And so it’s really an augmentation of what it is to be human and what it is that we’re able to do as humans rather than something which is a replacement for human beings.

Mills: So final question to wrap things up, what are you studying right now? What are the big questions that you’re trying to answer at the moment?

Griffiths: Well, a lot of the big questions that we’re focused on are really these questions about these distinctively human aspects of intelligence. So what I want to understand are these things that human beings do that are still challenges for AI systems, and these things that I think we could argue are sort of characteristic of what makes us human.

So we do a lot of work trying to understand how it is that humans are able to learn from small amounts of data and put those same kinds of abilities into machines. We do work, which is asking, how is it that people make efficient decisions about how to use their cognitive resources? And again, how that’s something that could potentially inform the kinds of algorithms that we use for solving problems like planning. And finally, how it is that humans interact with one another to achieve more than they’re able to achieve individually. And I think that last piece is particularly important in the world that we’re now entering where those interactions are not just interactions with human beings, but also interactions with other kinds of intelligent systems and trying to understand how we can structure those interactions and how it is that humans are able to make the most of our AI partners as we go into this future together.

Mills: Well, Dr. Griffiths, I want to thank you for joining me today. This has been really interesting and I appreciate your optimistic outlook.

Griffiths: Thank you. It’s been great.

Mills: You can find previous episodes of Speaking of Psychology on our website at www.speakingofpsychology.org or on Apple, Spotify, Stitcher, or wherever you get your podcasts, and please leave us a review. If you have comments or ideas for future podcasts you can email us at speakingofpsychology@apa.org. Speaking of Psychology is produced by Lea Winerman. Our sound editor is Chris Condayan.?

Thank you for listening. For the American Psychological Association, I’m Kim Mills.