Kim Mills: From writing emails to reading CAT scans, millions of people across a wide range of professions now use AI tools to help them do their jobs. In many cases, these tools save time and make us more productive, but as workers increasingly rely on AI tools for critical job tasks, some psychologists are asking what will happen to our own skills as we hand over more work to artificial intelligence?
People have worried about losing skills to technology before. We have GPS, so we rarely bother to look at maps anymore. Calculators change the way we do math. And in aviation, experts have long studied how autopilot affects pilots’ skills. So is AI just the latest example of this so-called cognitive offloading? How is it changing the way people learn, remember, and develop expertise? Are some skills more vulnerable to loss than others? Does it matter whether you’re a novice or already an expert when you start to use AI? And what can workers do to make sure they’re using AI in ways that help rather than hinder their learning and development?
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. Brooke Macnamara. Dr. Macnamara is an associate professor of psychological sciences at Purdue University, where she directs the Skill, Learning and Performance Laboratory. Her work explores how people acquire skills and expertise in professions ranging from elite sports to medicine. She looks at how practice, training, cognitive abilities, the environment, and other factors affect skill development. She’s currently studying how AI may affect learning and performance, and she recently spoke with APA’s magazine Monitor on Psychology for an article on the topic.
Dr. Macnamara, thank you for joining me today.
Brooke Macnamara, PhD: Thank you. I’m thrilled to be here.
Mills: I mentioned in the introduction that people have worried about losing skills to technology before. Can you talk about that? I mean, how have we offloaded certain tasks to technology and is AI in any way different from these previous technologies?
Macnamara: Sure. We’ve been offloading to technology, even things that we don’t think of as technology, for quite a long time. So back in 370 BCE, Plato lamented that we would lose our memories and those skills because we suddenly had writing and that people would then just start using writing and not work on the great memorization that a lot of people had where they would learn stories that could go on for days. And he’s not wrong, right? We did lose those skills, but we’re okay with them because most of the time we do have access to jotting down ideas.
And this is of course—we have gone through the times where we have lost some of these skills. Depending on the age of the listener, you might remember knowing a lot of phone numbers when you were a kid. And we don’t know too many phone numbers these days, but that’s something that we’re just sort of willing to be okay with that we’ve lost that. You mentioned GPS. A lot of people, at least anecdotally, don’t think that they have the same spatial navigation skills or will sort of develop a good mental map as quickly as they used to without GPS. But again, we usually have access to it. So these are all skills that have indeed, at least for a lot of people, maybe have reduced or atrophied, but generally as a society, we’re okay with it.
That said, whenever a new technology comes around, there is kind of this panic. So Orben defined it as the Sisyphean cycle of technology panics. And so we’re in a state now, so not too long ago it was the internet and Google. There was a famous article, is Google making us stupid? And we’re sort of living through the end of one cycle potentially and moving on into a new one.
And the issue is that technology moves pretty quickly and science can move less quickly, at least psychological experiments. So how this affects us, we can be a bit behind on knowing how much it affects us. But if we think of skills and knowledge acquisition as use it or lose it, or of course to learn a skill, we have to engage with it. If we are cognitively offloading to something else and we’re not engaging in the same way, then there’s a good chance that we’re not learning the skill as deeply as we would if we didn’t have that technology.
Mills: Do we know yet whether we’re losing skills to AI, especially in the workplace, or is it still too early to know?
Macnamara: Well, it’s certainly too early to know. We are in the stage where there’s early experiments that are starting to come out. Of course, we don’t know what will replicate, what will change as the technology changes, but there is some information out there that is potentially concerning. So for example, there was a study with endoscopists where at a hospital an AI tool was implemented where they had random access to it in terms of being able to detect adenomas, so polyps. And if you look at their detection rate before the AI was implemented and then randomly when it was taken away, their detection rates fell. So it suggests that they potentially lost some of that skill and how well they could detect without the AI available, that they’d gotten used to the AI being available.
There was another study that found that medical professional skills like basic life support decayed when people didn’t use them. And this was roughly decayed by about 50%, so 40 to 60% after 6 months of disuse. So if we are offloading to AI and not really using these skills to the same degree, there’s a good chance that those skills will decay.
Mills: So speaking of whether Google makes us stupid, I mean that’s an interesting article and actually dates to 2008. And of course the author wrote about this lack of attention span, that he couldn’t read a whole book anymore the way that he used to. And I’m just wondering if this is the kind of thing that we should fear with AI as well. I mean, are there things that we can do to keep those skills sharp?
Macnamara: Sure. So cognitive offloading isn’t kind of an on/off switch. So it depends on how we’re engaging with the AI tool and how much of the work we’re cognitively offloading. For example, if someone is trying to learn a new skill, but in that process they’re really just asking AI for the answer every time, they’ll get some exposure, they’ll probably learn a little bit, but not as much as if they were trying to engage in what are called desirable difficulties. So if they are trying to generate the answer themselves or test themselves, what’s called retrieval practice to try to remember the answers, they’ll learn the skill better that way. So if they completely offload, probably they won’t learn that new skill to the same degree as if they didn’t have the tool.
But there are other cases where you could use the AI to check your answer or give you examples, ways where you are interacting and engaging and you’re not just completely offloading all of the cognitive work—where you really are using it as a support system. In which case there’s evidence out there that suggests in some case you do better if you have that AI support, or in some cases you’re at least not doing as poorly as if you are completely cognitively offloading. So it really depends on how we’re using it.
Mills: Are there other tips and tricks? I mean, because some AI will basically not only give you the answer when you ask a question, then it will ask you if you want to know these five other things. I mean, maybe you as the user should be thinking of what you need and not relying on AI to tell you Here’s the next question you should ask me.
Macnamara: Right, right. So there’s always the question of, are we losing critical thinking skills by offloading that part to AI? And there’s evidence that maybe we are. Interestingly, since we were just talking about Google and people having concerns about it, now there’s been studies comparing Googling things to using AI. And in those cases, well, this is a study, I guess I should say, we don’t have a lot of data yet, but actually it was one paper, multiple studies where people were asked to learn about something and then give advice to a friend on the topic. So how to plant a vegetable garden, how to live a healthy lifestyle. And either they Googled information on a few pages and then synthesized it themselves and then gave that advice or they used AI, which of course synthesized it for them. And what they found is that for those who used AI, the advice they gave was briefer, it was less unique to the person, and it was rated as less helpful than if they used Google.
And the idea is that AI was sort of doing that cognitive work for you. You put in less effort and you weren’t having to synthesize that information yourself. You weren’t having to go in, read multiple tabs, think about how it all works together and then produce that advice. It was just given to you and therefore it was less interesting and the person didn’t seem to learn as much.
Mills: It sounds like the good old days of the World Book Encyclopedia where that was all we had and you would have to go in and read it and synthesize it yourself. So what you’re saying is doing a little bit more of the work yourself is advisable at this point if you want to stay sharp.
Macnamara: Certainly that’s some advice. Now of course, nobody is suggesting that you give up on computers and the internet and go back to the library and having to pull up each encyclopedia at a time. Nobody wants to live that way, because there’s a lot of benefits of course to technology. So Google is much faster. It is at your fingertips. So it’s more about how you’re engaging with the technology and how much you’re offloading. But yes, absolutely. I think the more that you engage with the skill yourself, test yourself on how well you know the information or can perform the task without AI is important because there’s also the possibility that we are overconfident in our skills with AI. So if you are engaging with the task with AI, probably your performance better than if you weren’t using AI. But if that AI is taken away, what is your skill level?
But because you’ve been performing the task at a very high level, it’s possible that you’re mistaking your competency for what is essentially your competency plus the AI’s competency. So it might be good also to test your skills without it to see what your actual competency is. And this is something done—so for example, autopilot, which of course is not AI, it’s an automation. In other words, if you have the same environment, it’ll do the same thing every time, which of course is different from AI, which is learning and adapting and making decisions that were previously the domain of the expert alone. But in aviation autopilot, the FAA is concerned that pilot’s skills are atrophying, that they’re not performing as well. And so there’s strong recommendations that pilots go through an entire flight using manual control because they’re concerned that those skills are being lost.
Mills: Well, now we’re talking a little bit about expertise here, and I know you studied that. How might using AI affect novices like medical residents differently from experts like experienced surgeons?
Macnamara: Right. There’s again, limited evidence, but some evidence to suggest that these are different processes. So this is outside of the workplace, but for example, there was a study looking at using AI to learn math. And so if you’re learning a new skill, or even with calculators going back even more basic, with calculators, if a kid is learning a new math skill, but with a calculator, they tend to not learn it as well as if they needed to try to generate the answer themselves and then use the calculator as a check. But for kids who already had those math skills and then used a calculator, it didn’t matter. It didn’t affect those skills. Now, long-term, if the study had looked for a very long time, would those skills have atrophied in the kids? Possibly.
So in the workplace, there’s the idea of, okay, what’s happening during training where you’re learning a new skill? And then once you’ve already hit a level of expertise, but now you’re just relying on the AI, do your skills decay? So there’s a bit of evidence for both of those. It’s likely the case that if you know a skill really well, it’s going to take much longer for the skill to decay to have that disuse. We have that study that I mentioned that after about 6 months of disuse that those skills start to decay by about 50%. Now that’s complete disuse. Using the AI and cognitively offloading isn’t complete disuse so it’s probably a bit slower than that, but I think it’s something that we still need to be concerned about.
Mills: If someone’s skills are starting to decline because they’re relying on AI, will they even notice it? I mean, is it obvious to people that they’re leaning on AI and they don’t really know what they’re doing themselves?
Macnamara: Quite possibly they would be completely unaware of it. It’s going to depend on how they’re using the AI most likely, but absolutely if the AI is picking up the slack, so to speak, at every turn and keeping that inside a black box so that they don’t even know that it’s happening, absolutely we could be caught unaware, which could potentially be an issue in say cases in medicine where not only maybe the AI is not available or gets updated and it’s different and you don’t know how to use it. So you’re a surgeon using a robot—and this is a bit more in the future because the AI isn’t really set in there, but that’s a goal in terms of the future of work. So now the AI isn’t working. Can you still perform the surgery? So there’s that aspect that we need to be concerned about.
Mills: I’m just wondering, since you’re studying AI in professions that include things like elite sports, how does AI influence somebody who’s an elite athlete? I mean, you still have to perform if that’s what you do. So what is AI doing for you and are you actually offloading anything to AI when you are an athlete?
Macnamara: There’s probably less that you’re offloading. There could be decisions that you’re making or learning or skills, but you’re going to essentially be forced to test yourself without AI present. As of now, there’s not AI in a racket for a tennis player to be told at the moment like, no, go for the spin. They’re about to do this. So I think for now at least probably that’s pretty safe as opposed to again, going back to medicine where not only maybe the AI is unavailable, but maybe the AI is available, but we get say a novel coronavirus. And so we don’t have the data needed for machine learning for AI to be able to make decisions. We want to make sure that our medical professionals can still reason and think through information in those cases where the AI isn’t going to be available for a different reason.
Mills: For listeners who are already using AI in the workplace, what should they keep in mind to make sure that they’re using it in a way that still keeps their skills sharp?
Macnamara: I think that it would be good to make decisions that are well-informed. So if this is a skill that you don’t care about, then don’t worry about it. So if it’s, for example, you want to spend your time coming up with a new business idea, you don’t want to program the budget, for example, come up with the budget. Yeah, maybe using AI for that while you spend your cognitive resources on something else, that might be a good use for you. So it depends on what skills you want to make sure that you maintain. If you want to learn a new skill or you want to make sure that you maintain a new skill, then really thinking about how you use the AI. And I think this is important in the workplace in particular because the short-term benefits are always going to lead to saying use the AI.
Because in the immediate, the short-term effects are it’s probably you’re going to get a better result. It’s probably going to be more efficient. It’s going to be less effortful. But long-term, if that’s a skill that you don’t want to lose, then maybe think about not using AI every time, testing your skills without AI, or using it in a way where you are generating first and then asking for feedback where you are interrogating the prompts that you’re giving it and the responses that it’s giving you, or asking it to fill in gaps that you’re missing. And the next time try to do it without the AI providing you those cues and see if you can do it.
Mills: What are you seeing out there as you do research? Are workplaces thinking along these lines as they implement AI procedures at jobs?
Macnamara: I think they are. I mean, I’m sure there’s a lot of variability and some places probably are thinking about it more or less. And of course, people are using AI in all sorts of very different ways from creative endeavors to more procedural endeavors. So for example, one way that it’s being used is in radiology. In radiology, a lot of systems either have just a computer-assisted device or it is true AI that’s helping find abnormalities in an image. Now it’s always the recommendation that radiologists first try to find the abnormalities themselves and then click the button to see what the system says it’s there. But again, we don’t really know what they’re doing. So if they really don’t feel like putting in the effort or they’re less concerned about their skills, they might just be clicking that button and then hopefully at least checking the work.
And again, immediate reward is that it’s less time. They’re going to be more efficient and probably have a high success rate in terms of detection. Long-term, we want to worry about their skills and making sure that they could still do this without the AI because AIs are not perfect. They are fallible and people are fallible. But if we start leaning too much on the AI or forgetting about its fallibility, then we’re not putting that shared cognition together. So we have the human intelligence, we have the artificial intelligence. And the idea is if we look first, we’re more likely to find something that maybe the AI is going to miss. But if we just look at what the AI says, we might not see what we would’ve caught before. So there’s that concern. And people very much differ in how much they are biased towards leaning on an AI versus not. So there’s some people who get complacent when an AI is available. There’s some work that suggests, and this is in aviation and automation in general, that once a system is about 70% accurate, people tend to just rely on it. And 70% is still leaving a lot on the table.
Mills: Some people argue that AI is going to free people from a lot of routine work, what I call the tricky track that you don’t want to do so that you can focus more on creative and difficult tasks. So do you think that AI is going to lead more to upskilling than to de-skilling?
Macnamara: Absolutely in some ways. It’s hard to say what that proportion or ratio is, what is going to be more upskilling compared to less de-skilling. Certainly it’s going to be the case that people will be able to not spend their time on the tedium and might be able to learn new skills that they’re more interested in. It’s also probably going to be the case that people will lose skills and as a society we might be okay with losing those skills. Which one is going to be greater is hard to know. We certainly hope as a society moving forward that it’s going to be greater the upskilling than the de-skilling, but it’s a bit early to tell.
Mills: Would you come down more on the optimistic side or the pessimistic side based on where we are right now?
Macnamara: Well, I think we’re going to end up doing both. Lots of things will improve as usually is the case when we have a new technology. We will certainly lose some skills that as a society we might be okay with losing or as an individual we might be okay losing. What’s interesting about AI is some people would say that the argument is it’s just another step in technology and it’s not really anything special and we’ll move forward in the way that we have with all of these other new technologies. Then others would argue that there is something different about AI that unlike other technologies, it’s taking components of our cognitive processes that were just not available before where it is making potentially decisions, expert level decisions that we wouldn’t have offloaded before that we are now. So there’s not only the de-skilling concern and the skill decay concern, but there is the overconfidence concern.
There’s a concern that maybe we will generally start expecting immediate results more, not wanting to persist as much. There’s been a little bit of work already in this space. This is from a pre-print, so of course you have to be careful with that because it hasn’t been peer reviewed yet. But across multiple studies, they noted that when they were trying to learn a new problem-solving skill and they used AI and they got the answer and then when it was taken away, people just gave up a lot more quickly. They got used to getting this immediate result and it just wasn’t worth it anymore to put forth the effort. And so there’s concerns that this could then spread and be broader. Now the research is sort of all over the place in terms of whether you try to encourage people to put forth more effort in a task, whether that has far transfer or whether people start persisting more generally.
So there’s some research that suggests yes, some says no, that’s not really the case. But the idea is whenever we think that something might improve with an intervention, well, then it might be the same thing that if we have this intervention, in this case AI, and we’re not using a cognitive process, then it might get worse. So kind of the opposite of what we usually think of in an intervention. So it’s hard to know whether it’ll change expectations or habits in terms of diligence and industriousness and persistence and things like that, which is a concern. And maybe that’s been a concern with past technologies and there’s nothing different here, but these are the current concerns that we’re thinking about.
Mills: Do you use AI in the work that you do? I mean, does it assist you in your research or check your math for you?
Macnamara: Oh, okay. So no, I don’t use AI at all. I mean, other than of course there’s parts that you can’t get away from. So now you use Google and an AI synopsis comes up, and so I will look at that and I do try to click on the link and actually double check it. I ignore the email that tries to, you’re responding to an email and it tries to finish your sentence for you. I generally say, Hey, stop that and do it myself. So no, I’m probably one of the few people studying AI and not using AI. Perhaps as somebody who studies skills and performance and offloading, I’m sort of just not willing to offload at this point. That might change. I might just be—
Mills: Resisting the inevitable.
Macnamara: Right, exactly. Exactly. So we’ll see. But I don’t want to come across as just sort of anti any kind of new technology because I think there are absolutely benefits that come with it.
Mills: What are you working on now? What are the big questions that you’re trying to answer?
Macnamara: So one of the questions that I’m working on with colleagues from computer science and medicine and robotics is looking at AI skill development, particularly in radiology and surgical robotics, skill maintenance in those areas. So do we see skill decay? And then overconfidence. So we’ve been collecting data on that for about a year. We’re getting close and it’s still preliminary at this case, so I don’t want to get into results, but it might be potentially the case that we should have some concerns about AI in terms of our confidence and our learning. So that’s something that I’ve been working on with AI. And then with Evan Risko, who studies cognitive offloading, we’re planning a series of studies to look more at some of these studies that have come out and looking at longer scales of skill acquisition, other ways that cognitive offloading might change how people interact with AI.
So just sort of joining what a lot of people are studying this topic at this point is just that there’s lots of new studies coming out, but even though I say lots, we’re so new. It’s just these one-off studies on this one topic here and there that we’re getting to where we really have a broader knowledge base about it, but we’re not there yet.
Mills: Anything you want our listeners to know that I didn’t get around to asking you?
Macnamara: Sure. One thing that Evan Risko and I have been talking about and some of his colleagues is the likely difference in effects of using AI on skills and knowledge versus basic cognitive abilities. So skills and knowledge are both acquired primarily through practice, through experience. There’s some kind of innate knowledge, but for the most part, when we acquire knowledge, we have to have some sort of experience where we learn it. And likewise with skills, we’re not born knowing how to play chess, for example. We need to practice that. Now it’s not that that is 100% of it. People vary in how quickly they learn, for example. So it’s not that practice explains 100% of our skills, but it does explain for the most part changes within a person. So if I practice the piano tomorrow, I’m probably going to be a little bit better than I was today and so on.
So there’s skills and knowledge are very sensitive to our experiences and therefore potentially how much we’re offloading. Basic cognitive abilities or fundamental cognitive abilities, foundational abilities—again, these are not perfect distinctions, and of course there’s bleeding into skills and abilities—but broadly speaking, abilities are more stable, your reasoning ability, your working memory capacity, how much information you hold in your mind. And of course there’s development as kids and of course there’s declines at older age. But broadly there have been a lot of interventions trying to improve basic cognitive abilities. And while at first some of those studies looked promising, usually it’s the case that they were unable to have far transfer. So we have people do a task that’s designed to increase their working memory capacity and then we see if they’re better on an intelligence test and it’s really hard to find that far transfer. So it suggests that we’re fairly insensitive to our experiences with those cognitive abilities.
So for a lot of people, that was bad news. For a lot of people, we wanted to get smarter by doing brain training and that seems like a great thing to do. The promising side of this, of course, is that if these abilities are fairly insensitive to experience, then we probably don’t have to worry as much about our intelligence getting worse with AI because if it’s insensitive, more insensitive to experiences, it’s probably not going to be the case that we use AI and now we just can’t think anymore. So skills and knowledge acquisition, much more sensitive to experience and probably much more sensitive to how we use AI in terms of maintaining and acquiring. But basic cognitive abilities we might be a little bit safer about.
Mills: Well, it sounds like we need to talk again in another year or two to see how things are going.
Macnamara: Absolutely. Yes. I’ll give the caveat that all of this is new. I could be completely wrong about everything that I’m saying.
Mills: Well, I doubt that, but I want to thank you for joining me, Dr. Macnamara. This has been very interesting. Thank you.
Macnamara: Thank you so much for having me.
Mills: You can find previous episodes of Speaking of Psychology on our website at speakingofpsychology.org or on Apple, Spotify, YouTube, or wherever you get your podcasts. And if you like what you've heard, please subscribe and leave a review. If you have comments or ideas for future episodes, you can email us at speakingofpsychology@apa.org Speaking of Psychology is produced by Lea Winerman.
Thank you for listening. For the American Psychological Association, I’m Kim Mills.
Brooke Macnamara, PhD, is an associate professor of psychological sciences at Purdue University, where she directs the Skill, Learning and Performance Laboratory. Her work explores how people acquire skills and expertise in professions ranging from elite sports to medicine, and how practice, training, cognitive abilities, the environment, and other factors affect skill development.
