The following transcript was generated using automated transcription software for the accessibility and convenience of our audience. While we strive for accuracy, the automated process may introduce errors, omissions, or misinterpretations. This transcript is intended as a helpful companion to the original audio and should not be considered a verbatim record. For the most accurate representation, please refer to the audio recording.
Michael Dunne: I’m Michael Dunne. You might think that very tech-savvy college students would be the most judicious and careful users of AI in their studies, because they are perhaps more knowledgeable about its pitfalls. But according to a new study from Oregon State University, you’d be wrong. Today on the show, you’ll hear from researchers at OSU about their study, which shows how STEM majors at several universities are becoming overreliant on AI, and it’s hurting their educational experience. One of the biggest problems with this overreliance is students offloading critical thinking skills to a machine instead of flexing their own cognitive muscles, and this may have truly profound and negative consequences for their future. Joining us are Rudrajit Choudhuri, a doctoral candidate at OSU; Anita Sarma, a professor of computer science; and Margaret Burnett, a professor of computer science at OSU. I really appreciate all of you joining us.
All: Thank you.
Dunne: Yeah, I read a report about your study, and it says tech-savvy STEM students are prone to letting AI do their thinking for them. It’s a very provocative statement. Rudrajit, why don’t you tell us what your study found?
Rudrajit “Rudy” Choudhuri: Yeah. Let me give a little bit of context about the study first. We ran this large-scale empirical study with students at five different North American universities to find out how trust in generative AI and routine use of generative AI affect their cognitive engagement in STEM coursework, and which students, with which cognitive styles, are more affected. What we found is that when students trust and routinely use AI, their critical thinking skills go down, their need for understanding goes down and their reflection goes down. Reflection went down by 66% and critical thinking by 41%. Even the need for understanding went down by 21%.
Margaret Burnett: Ah, this is with their coursework, right, Rudy?
Choudhuri: Yes.
Burnett: And interestingly, students who are touted as tech savvy, with traits that are celebrated in STEM, for example, risk tolerance, high self-efficacy and motivations to use technology, were more prone to this kind of cognitive disengagement in their coursework through repeated trust in and usage of Gen AI, by 15% to 30%, actually, in the scale of the effects. So maybe a way to put that is, your stereotypical computer nerd genius is the most likely to be negatively affected in how well they are thinking and absorbing and, yeah, poking around the edges of their technical coursework.
Dunne: Interesting. Yes, and that was professor Burnett who just weighed in there, for the audience.
Burnett: Yes. Sorry, I’ll add one more thing. Neither prior usage of Gen AI tools nor academic seniority protected these students from getting cognitively disengaged. Everyone across the board who participated in the study had these effects.
Dunne: OK, professor Sarma, I’m going to go to you. Tell our audience how the team conducted this study.
Anita Sarma: Cool. One other quick thing I want to add to the results we’ve been talking about before I go into the methods. One of the reasons we see that people who are using Gen AI more are thinking less, and that those who are more technophilic, those who like technology, are actually struggling with this, is overreliance. We talked about trust, and basically what happens is the more you trust the AI, the more you use it, the less you think. This creates a cognitive debt cycle because of overreliance, and it keeps going as a feedback loop. Going to the method, because generative AI often gives you the answer, we wanted to know how much Gen AI usage impacts our thinking. We used cognitive psychology literature to find survey questions about critical thinking, need for understanding and reflection. We also wanted to find out about these characteristics of overreliance, so we asked questions about how much they actually trust Gen AI. From there, we created a set of survey questions that are psychometrically validated, so they’re good questions that find you the results you’re looking for, and we used past literature. Then, as Rudrajit was saying, we provided the survey to students at five North American universities and collected their data. From there, we analyzed these factors of trust, adoption and usage, as well as demographic traits, like their AI experience and their engineering or undergraduate experience, and the other traits we talked about: technophilia, motivation, self-efficacy and risk aversion. We created a theoretical model to see the impact of these factors on this critical engagement, and from there came what Rudrajit was talking about earlier: the reduction in reflection, critical thinking and need for understanding that we found.
Dunne: OK, professor Burnett, I’m going to go to you. How do you define, and perhaps measure, overreliance? Because obviously, I imagine there might be a sweet spot in utilizing AI. How did your study look at this and say, OK, when you cross a certain threshold, that becomes overreliance, and then we get into problem areas? Or maybe that’s really hard to define. I leave it up to you.
Burnett: You know, I’m not even going to attempt to answer that question, because Rudy is the brains behind this operation, and Anita provides even more brains. The third real brain behind this whole operation is our other author, Christopher Sanchez, who is, I’ve forgotten his exact rank, but he’s on the faculty in the psychology department. He’s making sure we’re really toeing the line on the things we’re drawing from the human sciences as well as computer science, so I wanted to make that point. The other thing I wanted to talk about is how we found out about their cognitive styles. Maybe that’s a question you’ve got later on down the road. I can answer it now, or ...
Dunne: Go ahead. Why don’t you answer that?
Burnett: OK. So, in addition to those validated surveys from psychology that Anita just talked about, on reflection, critical thinking and, what was the third one, Anita?
Sarma: Need for understanding.
Burnett: Need for understanding. Yeah. Anyway, in addition to those, we also measured each individual’s placement on a set of spectra that we have on cognitive style types. These come from the GenderMag project, which is a project about inclusive design. There are five, which we’ve drawn from all kinds of human sciences, and they have been very well validated as well. Three of them are the ones Anita just mentioned. There’s technophilia, which in the GenderMag project is called Motivations: Why is somebody doing this technology stuff? Is it because they’re trying to get something done, or is it because they just love technology? The technophiles she talked about were on the “I just love technology” end. Let’s see, there’s also computer self-efficacy, which is their self-efficacy with technology itself, and their attitudes toward risk when using technology. Those are three of the five spectra of cognitive styles that GenderMag deals with. The GenderMag project also has a validated survey that has been used in a number of projects all over the world, so we used that to pick up people’s placement along these spectra, and that’s how we know what their tendencies were.
Dunne: OK, Rudrajit, I’m going to go to you. I’m interested in helping our audience understand. Part of this study introduces, at least to a layperson, these two groups: System 1 and System 2 thinking. Can you explain what that means and its impact on your study?
Choudhuri: Yeah. This is actually a concept that was introduced by Daniel Kahneman. He is a Nobel laureate, and in his book called, what is it? I forget the book name, actually.
Burnett: “Thinking, Fast and Slow.”
Choudhuri: Yes. Think of System 1 and System 2 as the two modes of how humans think. When I say 2 times 2, your brain immediately autocompletes it to 4, right? Or when I say “bread and,” your brain autocompletes it to “butter,” right? But when I say 17 times 24, for example, your brain goes into an algorithmic mode and tries to process and solve it. System 1 is basically your fast thinking, or your heuristic-based thinking. When we are talking about overreliance, what it’s training our brains to do is, whenever you see a problem, “Let me ask Gen AI,” right? That becomes your default autocomplete whenever you see a problem. Whereas earlier, when people built things through practice, when I said 17 times 24, you actually thought about solving it. You actually engaged with the problem. What we hypothesized, and then found in our study, is that it’s not just one-time usage. If you routinely use these tools, it calibrates you so that whenever there’s a problem, it’s “Let me go and ask Gen AI.” That’s the loop that causes this underlying cognitive debt cycle.
Sarma: I like to put it this way. What Rudrajit was describing is cognitive offloading. Instead of thinking through 17 multiplied by 24, or whatever number he said, you think, “Actually, it’s too hard to think. Let me cognitively offload it to my Gen AI, or my calculator, or the GPS for the next direction.” Basically, our brain is also a muscle. The less you think, the lazier it gets. Your cognitive musculature kind of goes down. And when you really need to think, when you don’t have a calculator or Gen AI, or the GPS is missing and you have to actually think, and you’re used to this System 1 thinking and not doing deep thinking, it’s really, really hard at that point to get your cognitive muscles back into gear. The more we offload, the more we start relying on these tools. Sometimes we call them cognitive crutches, and that’s what we need to save ourselves from: just offloading and relying on them. Yeah. The more we do that, the more we start overrelying on them, and then when the GPS doesn’t work, you’re driving off a half-built bridge because your GPS told you to.
Burnett: Yeah, I’d like to add a little more to that too. We’re not the first ones to notice this cognitive offloading. There’s a bunch of literature by other people showing short-term effects like this. The most famous one, perhaps, is the MIT study that had Pattie Maes as one of the authors. You may have run across it before. They actually had sensors on the brain, watching people solve some kind of problem using whatever their skills were, while another group had Gen AI available, and for that group the brain is just kind of lying there dead, not doing anything, whereas for the other group it was very active. But what makes our study different is that it’s longitudinal. It’s talking about routine use of generative AI in their coursework, so it’s very situated in their real-world lives, and over time. I think ours is the first to talk about that in an empirical, quantitative way, and to actually model the impact of Gen AI usage on critical thinking and reflection.
Dunne: As I understand it, Rudrajit, and this is a question for you, in the study you all make suggestions for people to improve their cognitive skills. Maybe talk a little bit about those suggestions.
Choudhuri: Yeah, what we suggest is along the lines of breaking this cognitive debt cycle, which is a cycle where you rely more, you use more and you cognitively offload more, and soon enough you have this cognitive surrender where you’re like, “Why bother thinking? Let’s offload to Gen AI.” We suspect that it’s not just because of the tool that people are doing this. For example, think about when you’re reading a really good book that you like to read. You won’t be intrinsically motivated to offload it to Gen AI, because that’s the fun of reading. That’s something you are intrinsically motivated to do. Going back to Kahneman and these psychological theories, they suggest that human minds are neither lazy nor super hardworking. They are rational, so they respond to the environment they are in. For example, in STEM coursework, students might look at the assignment, or maybe the curriculum or what they’re reading, and think, “OK, why should I have to memorize this assembly-level language, or why should I have to code in Python, when I can just offload it to Gen AI because Gen AI can do it well?” Now I want you to realize the mismatch here. Students are thinking, “OK, the things I’m learning in school are no longer relevant because I can just offload them to Gen AI.” Meanwhile, instructors are like, “Hey, this is for you to build your critical thinking skills and problem-solving skills. I don’t care if this level of abstraction goes away tomorrow and you’re not writing code anymore. This is to teach you problem-solving.” So I think there is an opportunity there to intervene by making coursework more cognitively engaging, where you play with these cost-benefit levers. One is social accountability. We have seen that when people work in groups, their motivation to just offload to Gen AI and do a lazy job is reduced. There’s literature that shows that. Another way is gamified environments: when you have puzzles, when you have quizzes, when you have things that intrinsically motivate people toward a reward structure. There are, of course, pros and cons to it, but that’s also a way to raise the benefit lever. Similarly, in generative AI design itself, you can tweak these cost-benefit levers to have metacognitive scaffolds that, instead of giving the answer away, teach you: “Hey, have you thought about this? Have you considered this approach?” That not only lets students use AI better but also trains their cognitive musculature while they’re doing so. Those are some recommendations that we have.
Burnett: And maybe I could add one other thing to that too.
Dunne: Sure, professor Burnett.
Burnett: One of the implications of what Rudy just said is that it’s not inevitable that AI be designed the way it’s currently designed. It could be designed differently, in ways that wouldn’t cause people to use it as a crutch instead of thinking. And one other thing: This is particularly true of students. Once students have missed the fundamentals and later realize, “Whoa, that was a bad idea. Now I don’t have anything to build on in these new courses I’m in, and Gen AI doesn’t know the answers to these questions,” well, they’re kind of sort of dead, because they don’t have the fundamentals they need. Now they’re stuck with Gen AI. If they want to pass their classes, they’ve got no recourse but to keep trying to eke out answers from it, even when the answers aren’t there.
Dunne: OK, professor Sarma, I’m going to give the last word to you, because we’re running up against time here. Is there a correct diet, if you will, of utilizing AI that works for most people? You don’t want to, again to use the analogy, overeat on AI, but there is also a proper level of utilization to be able to get by in today’s modern world.
Sarma: It’s a great question, and the right diet is slowly changing as models get different and better capabilities, and as what we do at work also changes. I think one of the things we want to keep human is critical thinking and judgment, because any of these Gen AI tools can give you answers and solutions, but which solution is actually right? Being able to compare, contrast and critically think about the solution space needs human judgment. Toward that, one of the things our group, at least, is looking at is how we can bring in critical thinking. We are using a framework called the Paul-Elder critical thinking framework to help people do intent articulation. The easier way to put it is: Think before you ink. You have to actually think about the problem space before you go to Gen AI, and then you can have this diet of Gen AI, if you have thought through the problem. Because if you have not actually thought through the problem, when the solution comes, it looks plausibly right, because you have not really thought about all the edge cases and parts of it. So I believe it’s not how much AI to use, but how you use AI, by thinking ahead about the problem, thinking about what it is you really want to ask for, thinking about the pros and cons, and once you get the solution, not just clicking yes. Think back on whether the solution is right. When I’m teaching classes, I always have my students use more than one model, ask a what-if question or ask a why-this-solution question, and then you will learn more and actually be using your judgment.
Dunne: This is fascinating stuff. I really want to thank all three of you for coming on. We’ve been speaking with Ph.D. candidate Rudrajit Choudhuri and with professors Margaret Burnett and Anita Sarma. They’re both computer science professors at OSU. Really appreciate all three of you coming on.
All: Thank you.
Dunne: That’s the show for today. All episodes of “Oregon on the Record” are available as a podcast at klcc.org. Monday on the show: While Oregon is a true-blue state, it wasn’t always that way. In fact, one man helped usher in the state’s Democratic leanings, and we’ll talk with my colleagues at OPB about their new documentary about this man and his legacy. I’m Michael Dunne, host of “Oregon on the Record.” Thanks for listening.
