A recent study shows that people using AI to write for them experience some negative cognitive effects. Why? Because there’s something special about what writing does in your brain.
Guests
Nataliya Kos-Myna, research scientist with the MIT Media Lab. Lead researcher on “Your Brain on ChatGPT: Accumulation of Cognitive Debt when Using an AI Assistant for Essay Writing Task.”
Barry Gordon, director of the cognitive neurology/neuropsychology division at Johns Hopkins University.
Also Featured
Audrey van der Meer, professor in the department of psychology at Norwegian University of Science and Technology.
Transcript
Part I
Did you know that one of the greatest philosophers of all time, Socrates, was actually against writing? In Plato’s works, he expressed some wild concerns about the written word. His main worry was that writing could actually harm our memory and understanding.
MEGHNA CHAKRABARTI: That’s a clip from the pop science YouTube channel, Talking Heads.
Socrates is of course, the ancient Greek philosopher who had a profound impact on classical antiquity and Western philosophy. And it is true that Socrates himself wrote nothing. He instead gave us the gift of the Socratic dialogue, where through conversation and questioning, we explore a topic, challenge assumptions, and ideally sharpen our critical thinking skills.
Socrates did not believe that could happen through writing. Writing, he thought, would make people lazy, forgetful, and inattentive. Now, arguments in that same vein have been made about typing, computers in general and most definitely the smartphone. And now comes AI and AI writing tools.
Socrates opposed writing because he said it was bad for memory. And people today oppose AI because of critical thinking. But if you look at things in a box, of course you’re going to have a minimizing and a maximizing, but in the middle, you’re going to have this growth to work through.
CHAKRABARTI: Almost upon arrival, ChatGPT and other large language models have changed how many people approach the act of writing.
The impact has been so fast, it’s already a point of popular discussion. Socrates would be proud, just like the one you heard from marketing consultant Hanna Laikin. Now, with any rapid technological change, it is worth examining how that technology changes, not only how we interact with the world, but how we process the world internally, and in this case, how we write and how we think.
Earlier this summer, researchers at the MIT Media Lab published a paper titled “Your Brain on ChatGPT: Accumulation of Cognitive Debt when Using an AI Assistant for Essay Writing Tasks.” And the top line findings were maybe not so great for fans of ChatGPT, Claude, or other AI tools. And joining me now is Nataliya Kos-Myna.
She’s senior researcher at the MIT Media Lab and one of the authors of Your Brain on ChatGPT. Nataliya, welcome to On Point.
NATALIYA KOS-MYNA: Hi. Thanks for having me.
CHAKRABARTI: So what was your first inspiration for wanting to do this study?
KOS-MYNA: When ChatGPT was released barely three years ago to date, actually, I started seeing at the lab, my students copy-pasting some texts from ChatGPT and I was like, what is exactly happening? And then honestly, collectively at the lab, we started to get a lot of emails from people. And of course we don’t know if those are real people, there might be bots, but a lot of people started saying, Hey, my memory’s changing.
I think this technology has some effect on cognition, and I am actually working on something called brain-computer interfaces. So we have all of this relevant equipment, so to say, at the lab, and I decided let’s try to understand what exactly is happening in one’s brain when you use a large language model like ChatGPT.
CHAKRABARTI: Okay. Interesting. So you were seeing something in your own students. This is fascinating. Now, you used students in order to do this experiment. So describe to us what you did.
KOS-MYNA: Yeah, absolutely. Because we actually do have obvious on campus access to students. So we decided, hey, let’s run a study. Invite, in our case, 54 students from Greater Boston area, schools like MIT, Harvard, Wellesley, Tufts to come in person to the lab. And actually, gives them a very specific task, which is essay writing. We decided to ensure that the task is something they can manage within pretty short period of time.
It was just 20 minutes, and to make things a bit easier on them, we decided not to give very specific topics which require a depth of knowledge, like expert knowledge. Like for example, I would ask you right now, oh, can you write me an essay about brain computer interfaces? Never heard about those.
It might be actually complicated. Even if I gave you two hours or 20 days instead of 20 minutes.
CHAKRABARTI: A couple of years probably, I had to be able to start.
KOS-MYNA: Oh, right now? Yeah, maybe. But regardless, so we decided to give them SAT type topics, like what is the perfect society? Do you need to think before you speak? Happiness and things like that.
And we divided those students in three groups. One group was allowed to use ChatGPT, and we call it ChatGPT in the paper. And you might already guess this group was only allowed to use ChatGPT. They couldn’t go to any other chatbot or anywhere else for any help. The second group was allowed to use search engine and more specifically, Google.
So they were allowed to go anywhere online, whichever site they prefer, but they were not allowed to go on any type of chatbot powered websites or services like ChatGPT, or Gemini, et cetera. Okay. The third group, you might already guess what happened there. We actually did not allow that third group of students to use anything except their own brain.
So they couldn’t use any chatbot, they couldn’t use any type of help, neither search, nothing actually was available to them except their own brain. So they were all in the same condition of writing an ass essay and we measured their brain activity. We measured it non-invasively, so we put caps on the head with electrodes.
This imaging technique is called electroencephalography or EEG. Some people might have actually experienced this in other studies or in medical setup. Then we also measured the outputs, of course, the essays themselves, so we analyzed what actually they have written. Then we asked two teachers, two English teachers after the study already took place, to help us score those essays.
So they didn’t really know anything about the studies. They’re not even from Boston. We just gave them these essays. We told them, Hey, these folks had 20 minutes. No one is English literature major. Please go ahead and give us the scores and just get back to us and explain how you did that. And finally, as it’s very common in these types of studies, we also asked our students several questions after each essay was given in.
For example, can you quote what you just wrote? And why did you write about this topic? Things like that.
CHAKRABARTI: Okay, so let me just jump in here. So you grade, essentially, you graded the essays for the quality of the argument, and you measured brain activity during the process of writing, and then you tried to gauge some benchmarks after, with the questions.
Okay. So I actually do want to dive into all three of those things in just a second. But jump to the conclusion. Nataliya, what did you find about the differences in quote-unquote thinking with your three groups here.
KOS-MYNA: Yeah, absolutely. So this by the way, was just the first part of study.
There’s also a second part of study, which a lot of people actually not paying attention to. But first part of the study, what we found is that if we are looking at the brain activity, so what we measured is specifically brain connectivity. And this is a very specific measure, that if we were actually in person together, you and me, and I would just turn to you, tell you something, you would turn back to me.
Tell me something. So we are measuring actually directionality, right? So who talks to who in the brain in this case and how much talking is happening? Did I just tell you, Hi, how are you? Or did I actually start talking? So that’s what we measured in the brain, not like brain rot or laziness or things like that.
Really how much talking is happening in the brain and where this talking is happening. So we actually found that with introduction of the tools, there is reduction of this brain connectivity. So to summarize it, for the brain only group, so students who were in brain only group, they showed the most widespread brain connectivity.
So we basically see this connection, this talking happening all over the brain. So a lot of this chart was happening like on front of your head, to the back of your head, from one hemisphere to another hemisphere. So all over the place. And guess what? Not surprising. Imagine right now I’m going to take away all of your tools that you have in the studio, and I’m going to tell you, now tell me everything you actually learned about this study.
And you cannot go online. You cannot check anything. You might not be even able to really properly spell my name right away, right? You would need to remember all of these details. So your brain is really like on fire, right? You need to work through memory. What you learned, what you heard, what you edit, actually told you about the study.
All of these things. And then when you look at Google search, so there is actually less of this connectivity, but it is very well present. And again, think about when you were online last time, right? You need to open a lot of tabs. You would actually incorporate information from one tab to another tab.
And this is, you see a lot of this activity specifically for those who are interested in listening to us in the back of the head. That area is called occipital cortex or visual cortex. A lot of that happening. There’s actually a very interesting paper about this called Your Brain on Google.
… So yeah, it’s back to the point that, hey, we actually do study these tools a lot in trying to understand what’s happening to the brain. And, but back to the result. ChatGPT Group, our final group, right, however, showed the least of this brain connectivity. And again, it doesn’t mean that the brain, as I saw in some media reports, went on vacation or stopped working.
The brain was working; it just was not as active in some of those specific areas. So there will a lot less of this chatter and exchange and conversation happening.
However, it’s very important to note that these results also extended to other measures that we actually performed in this study.
CHAKRABARTI: Other measures. Okay. But first of all, before we get to that, We have about a minute before our first break. Nataliya, let me ask you though the first question that I have, even just when you laid out the three outcomes in terms of just brain activity. Are you saying that we believe that higher levels of activity in those parts of the brain, map to what? Better thinking, better memory. Like what? What is the significance of that? Why do we want that?
KOS-MYNA: Yes. Why do we want that? Because effectively, based on the prior research, prior literature. It was established actually long before us, these areas of the brain are responsible for creative thinking, critical thinking, episodic memory, and very important language development.
So we actually do want to understand what it effectively means if you have less of that activity present, specifically in this case, connectivity.
Part II
CHAKRABARTI: Nataliya, you were describing your findings in terms of the different amounts of brain activity and the location of the brain activity with your three groups.
And just to sum that up again, that was the people who had to write an essay on their own, with no assistance. The group that was allowed to use Google, but no AI help. And then the group that was allowed to only use ChatGPT. So I understand that the brain activity map essentially was quite different in those three groups.
Now were you going to tell us a little bit more about the sort of the other half of this, the first part of your study in terms of the actual writing itself? Because you did have those, their essays graded. Were some better than others?
KOS-MYNA: Yes, absolutely. So first of all, what we actually found that while we performed analysis on those essays using NLP or Natural Language Processing, is that the essays of ChatGPT participants were very homogenous.
What it means is that they used very similar vocabulary.
The essays of ChatGPT participants were very homogenous.
What it means is that they used very similar vocabulary.
I’m going to give you a very specific example. There was a topic of happiness and what we found is that ChatGPT participants actually used a lot of vocabulary related to career and career choice. And you’ll be like, okay, great. Those were students, didn’t you say that?
So careers might be very important for them, right? For happiness. And actually, who are we to judge what is important for a person related to their happiness? However, those same students, right? They also used a search engine and then brain only group, right? For search engine, they actually showed a lot of vocabulary related to giving and for brain only group, we saw a lot of vocabulary.
So words that were used in those essays related to true happiness and happiness. This is just one of the examples, how different the essays were, the words that were used in the essays.
CHAKRABARTI: That is so interesting. So you’re saying that with each additional layer of technological assistance, let me just roughly call it, that you saw more and more homogeneity in the thinking that was on the page.
KOS-MYNA: Yes. Like literally the vocabulary that was used was the same, right? It was so different. And again, in the paper, we actually give examples for all of the topics that were used. And then finally what we saw, so from the teachers, right? Because then as I mentioned, they were effectively scores of the teachers, right?
They actually asked us several times, oh, were those two people sitting next to each other? We’re like, Oh no. It is like a personal setup. No way. They were sitting next to each other. They really marked themselves without knowing that these were ChatGPT essays and those were really coming from, most likely from an LLM.
And they actually were able, very interestingly, pick up micro differences in the essays related to a person. So what I mean here is that if you, for example, Meghna, you were actually our participant. You were not, but if you were, and then we gave to those teachers four, because there was actually the four sessions.
We’re gonna talk about it, in a bit. There were four essays, right, of yours. And we didn’t actually tell you that, Hey, this is participant one’s, whatever, essay, teachers were able to pick up these micro differences and tell us, oh, this seems like these four essays were written by this one person.
So in your case, by, like, Meghna. And it was very interesting because an AI model was not able to pick up on these micro differences.
CHAKRABARTI: Wow. Okay. You’re right, Nataliya. We’re gonna talk about the second half of your research as we continue to go through this hour, but you’ve set us up perfectly to bring in our next guest.
So that is Barry Gordon. He’s the director of the Cognitive Neurology and Neuropsychology Division at Johns Hopkins University. Barry Gordon. Welcome to On Point.
BARRY GORDON: Meghna, welcome. Good to see you. Talk to you.
CHAKRABARTI: (LAUGHS) We’re seeing each other in the theater of our minds.
GORDON: Exactly. You got it.
CHAKRABARTI: So let me first ask you with Nataliya’s research as the backdrop here when we’re talking about AI and specifically ChatGPT type tools to generate writing, right?
Because it can do all sorts of things. It can do calculations, it can do various analysis, it can do math, whatever. But we’re focused on writing, what is the most important? Or what are the most important types of quote-unquote thinking or brain activity you think that we should be focused on to understand the impact of AI on us?
GORDON: It’s a complicated question. Because it depends on what product you want, or the person wants. But I would say from the standpoint of the brain, if you’re generating something, whether it’s writing by pen, writing by typing, or whatever, if you’re thinking, that’s what’s most important. Because that’s what embeds it in your mind, and that’s what leads to the deepest possible processing and the best learning for many reasons.
If you’re generating something, whether it’s writing by pen, writing by typing … if you’re thinking, that’s what’s most important. Because … that’s what leads to the deepest possible processing and the best learning.
Barry Gordon
So that’s what I’d conclude.
CHAKRABARTI: So wait, no, explain a little bit more. When you said, do you actually mean writing by hand?
GORDON: Writing by hand turns out to be different than typing. Printing turns out to be different than cursive writing, at least in terms of the brain. And they all, first of all, we’re using, we’re ideally using our whole brain whenever we’re thinking.
And whatever tool you use to help you with that, so long as it’s your mind doing some of the work or a lot of the work, that’s what gets it embedded in your head.
CHAKRABARTI: I got it. Mental effort. Is it?
GORDON: Mental effort.
CHAKRABARTI: Okay. So let me refine my question. Thank you for saying that it was too big and too complicated.
What exactly is the brain doing when we are writing?
GORDON: (LAUGHS) First of all, hopefully you’re thinking about what you’re going to say and hopefully you’re taking time to think about it. Because, you know, that’s what people do more when they’re writing perhaps than when they’re speaking. And then, what are the actual mechanics?
What words are you selecting? For example, Madame Bovary, the author of Madam Bovary, was famous for having, Flaubert, was famous for having spent the day trying to think of the right word to use in his novel. That’s a lot of work and that shows in the brain. The brain in many ways is like a muscle.
Actually, many of the same things that work with training, work with the brain, or seem to work with the brain. Namely, you have to use it or lose it. You have to have it not too easy, not too hard, but in the right medium, or middle, to do it, to build up strength. And you have to spread it around, use a lot of different muscles so that everything is engaged together.
The brain in many ways is like a muscle. Actually, many of the same things that work with training, work with the brain.
Barry Gordon
And when you do that, whether it’s by printing, typing, writing, handwriting, or for that matter, even dictating your thoughts, you’re engaging lots of different processes, you’re engaging your general thinking, you’re engaging the word choice, you’re engaging how is this going to play? Is it going to make my point?
Does it mean anything? You’re going back and forth in your head. And that back and forth, all those things are involved. Everywhere around your mind.
CHAKRABARTI: So from what I glean from Nataliya’s research, and Nataliya, I’ll come back to you in just a second here, but that when more of the brain was engaged, when various parts of the brain, from front to back, hemisphere to hemisphere was engaged.
That indicated, I don’t know if I can say more effort or at least greater activity that led to some kind of better product or greater learning. Now I’m just inferring from Nataliya’s research. But is that, do we have reason to believe that’s true? That when more of the brain is engaged in a task, in this case, writing, that people learn it better?
GORDON: We definitely have evidence for that. The effort issue is a complicated one because as someone becomes more expert at something, typically the effort they have to expend diminishes, decreases. But on the other hand, since even experts are always pushing themselves or hopefully always pushing themselves, you’re always keeping up the effort somehow, even if you’re starting at a different level than what you began at before.
So you know, if you start lifting weight and you start with 10 pounds and then maybe it’s heavy enough for you, but eventually you might get to 100 pounds, the effort keeps building, relative effort, keeps building, and you keep getting better and better.
Her research helps confirm … is compatible with what’s been known about how the brain learns for quite a long time. Again, the more you put into it up to a point, the more effort you put into it, the more the brain makes those connections. Pulls out things that hadn’t thought of before and strengthens the weaving between them all.
CHAKRABARTI: I understand that we may have some complimentary evidence from people who are undergoing rehabilitation after having a stroke?
GORDON: Yes. So on the one hand, it’s fairly plausible and reasonable to imagine that the more parts of your brain you use, compared to less parts, you’re going to remember something better in general. Because you’re going to have it in more places, more active.
But there’s also some evidence that the brain seems to pay attention to how engaged it is. And for example, in stroke rehabilitation, if someone’s trying to just rehabilitate one arm, it’s not as effective as if the same motion is part of a whole-body activity or involving more than the arm. So it’s almost as if the brain is taking account of how important something is and saying, oh, this must be important.
I’m going to learn everything better.
CHAKRABARTI: Oh, interesting. So just to be clear, even if the target is still the same, the arm, there’s better outcomes if the whole body is engaged.
GORDON: That seems to be the case. I say seems to be the case. Because we all know medical research and research in general is often hard to replicate, but there’s that presumptive evidence, and it makes sense, too, that the brain wouldn’t waste its time on, if you’re just doing a single finger exercise, why bother?
But if that finger exercise is part of a whole limb motion, or your whole body playing tennis or something of that sort, then the brain will, in a sense, sit up and take notice and say, I better learn all of this.
CHAKRABARTI: Interesting. Okay, so let me go back to you Nataliya. What do you think about that?
KOS-MYNA: No, I absolutely agree with Dr. Gordon. And there is a lot of complimentary papers. I saw research, prior research, right? The earlier point that he made about this widespread connectivity, right? That what we measured in handwriting, right? There are several papers about handwriting that show that brain eases, again, showing this overspread connectivity can related to handwriting compared to typing, for example, et cetera, et cetera.
So that’s a whole idea of the fact that, hey, you are using more of the brain, of these different parts, so to say, no, this is very crude explanation. And I think what is very important back to the point, effectively expert level, is would actually be very important. But I think to put into perspective with our study on the use of AI, what I haven’t told you is the final finding —
CHAKRABARTI: … Well, actually, Nataliya, before you do that, I do want to, I promise you, we will hear the final. I’m just like teasing, I’m teasing listeners today, but since both of you have actually mentioned handwriting, I do wanna just pause for a second and talk about that.
Because in my personal N=1 of my life, I type every day like any normal person, almost all my writing is done through typing, but I don’t know, I find myself that when I like, really want to remember something or it’s of high value to me, I pick up a pencil or a pen.
And somehow that embeds it more deeply into my brain, specifically my aging, working memory.
So because of that, we actually reached out to one of the researchers that you mentioned who has done research on handwriting to understand like how we process information and think differently depending on how we’re physically writing. And according to on Audrey van der Meer the answer is quite a bit.
AUDREY VAN DER MEER: We actually looked at the connectivity patterns in the brain during handwriting and typewriting, and we found that the brain is much more connected during handwriting than it is during typewriting.
The brain is much more connected during handwriting than it is during typewriting.
CHAKRABARTI: van der Meer is a researcher and professor of neuropsychology at the Norwegian University of Science and Technology.
And in her research she used students, once again, the universal Guinea pig, as subjects. And it sounds like the experiment might’ve looked exactly how you’d think, but maybe with the fun twist.
VAN DER MEER: They were wearing one of our characteristic electrode nets. Consisting of 256 sensitive electrodes sewn together as like a hair net on their heads, and they were playing the game Pictionary. So we presented Pictionary words on a large screen, and they were supposed to either write the word by hand, draw the word by hand, or type the word on a keyboard. And we recorded their ongoing brain activity while they were performing those tasks.
CHAKRABARTI: And just like Nataliya’s team did, van der Meer recorded that brain activity on an EEG, and she says the number of brain functions, visual processing, sensory motor integration, and the motor cortex are notably more engaged when writing by hand.
VAN DER MEER: The brain does this through neural oscillations that can oscillate at different frequencies and in different parts of the brain and that kind, those kind of oscillations in the regions of the brain where we found activity are usually involved in learning and memory. So these kind of oscillations, they put the brain in a kind of state that makes it easier to learn from your handwriting activities and to remember what the notes were about.
CHAKRABARTI: Writing by hand can also promote recall of experience in place because it helps cement memories that are unique to us. According to van der Meer:
VAN DER MEER: Handwritten notes are very personal. When you then take them up in order to study for the exam, for instance, everything is coming back because it feels like you’re back in the lecture theater again.
And that’s why sending your mate to a lecture to take notes doesn’t work because notes, hundred of notes are typically very personal.
CHAKRABARTI: She says that personal also translates to the cultural.
VAN DER MEER: handwriting and drawing are part of our cultural heritage. And it would be a shame if the next generation would not be able to write a love letter by hand or even a poem by hand.
CHAKRABARTI: So that’s Audrey van der Meer. She’s a researcher and professor of neuropsychology at the Norwegian University of Science and Technology. Barry Gordon. We have one minute before our next break here and it’s hard for me to escape at least a partial conclusion that depending, and this is the big asterisk, depending on how you use AI or these large language models, that it really is engaging less of your brain and making us maybe think less creatively.
GORDON: People can use it or not use it that way if you like. But other people may use it to help expand their thinking, like every tool throughout history. We talk about writing, but remember writing wasn’t even around more than 5,000 years ago, so it’s a new tool we’ll have to get adapted to.
We’ve only had a couple of years with generative AI. Let’s give it a little bit longer.
Part III
CHAKRABARTI: Alright, back to our conversation about writing and AI. Nataliya, you’ve been wanting to tell us about part four of your research and now’s the time. Because it’s actually really important. So go ahead and tell us what you found.
KOS-MYNA: Yeah. Mostly just because a lot of media forgot about that part.
CHAKRABARTI: You say media only wants to focus on doom and gloom.
No, never.
KOS-MYNA: Yes. Not a quick bite at all. I know. No. But yeah, seriously speaking, and this is just immediately a pre-phase for that final phase of sessions. Four in the paper. We only had 18 participants coming back, unfortunately. So definitely the smaller sample size, but we still carried on with our study and we actually swapped the participants group.
So what it means is that if Meghna, you were our participant, and for regional three first sessions, you were a ChatGPT assigned group. For this fourth one, we actually would take access away and you would become a brain only participant. And vice versa was also correct. If you were originally from in our study for the first sessions as a brain only participant, for this fourth session, we’d actually get you access to ChatGPT, and we actually would do one more thing. We would get you not the new essays to choose from, to write your essay about, but we would get you the ones you had already written about in your previous three sessions.
So you already know them, you know the topics, they’re not new to you, so it’s not a surprise, an additional cognitive load, hopefully. So that’s what we did for this fourth session, and the results were actually pretty interesting.
CHAKRABARTI: Go on.
KOS-MYNA: So if you were originally a ChatGPT participant and then you became a brain only participant, your brain connectivity actually never matched the brain only participants, the truly brain only participants.
It was always inferior to them. However, if you were originally a brain only participant and then we gave you access to ChatGPT, your brain connectivity was significantly higher compared to the brain only participants. Meaning that if you did the work, so to say, with your brain only first, and then got access to the tool, you actually did better brain connectivity.
CHAKRABARTI: Ah, okay. No, keep going. Keep going. Go ahead.
KOS-MYNA: Yeah. Yeah. Of course, it’s only 18 participants, right? It’s very specific task. So we cannot draw any strong conclusions or generalize for everyone, but what it means that potentially timing might be very critical here. So when you’re introducing these tools and have you actually developed expertise, right, prior?
To using a tool in whatever domain, of course, we are trying to use AI in.
CHAKRABARTI: Okay. Let me ask you, in the first part of this study, in the ChatGPT only group, certainly you were monitoring how they were using ChatGPT, right?
KOS-MYNA: Yeah, of course.
CHAKRABARTI: And was it, did you find that there was various levels of did some people just put, here’s the prompt, write me an essay?
Or did other students say here’s the prompt, then they wrote an essay, and they said then ChatGPT wrote the first draft. And then they said no, fix this, fix that. Because I think even engaging with the tool, there has to be a great deal of variance there.
KOS-MYNA: Absolutely. You’re absolutely right.
We actually did monitor that and these results in detail actually presented in the paper, but to just give you and the listeners a bit of the overview, what is happening. What we found is that compared to this fourth session, so brain only to ChatGPT transition. For the original ChatGPT group, more than half did in fact ask, right, to just write the essay. Not everyone, right? Very important. Not everyone, some actually did provide some of their original draft. Some actually asked for specific questions, right? Hey, what is actually, is there any series about like perfect societies, right? Tell me more. What is there? Is it philosophical?
Is it economical based? What is it about it? But we found that for these four sessions that ChatGPT use after brain only use, actually, there was a significant reduction, two times less request for just write me an essay, and four times more requests. Just tell me more. [Instead of], write me an essay.
Just, Hey, tell me more about this topic. Tell me more about that topic. Give me that reference. So more what we call in the paper information seeking and some questions, searching precision or additional data compared to just prompt, direct copy, paste, prompt of the essay.
CHAKRABARTI: Got it. In a sense it’s a kind of Socratic dialogue.
If I can go back to the beginning of our show with ChatGPT. Okay. So Barry Gordon respond to what Nataliya has found there that, again, we can’t make any like hard and fast conclusions, but it seems as if when someone does that mental struggle first. And forms a kind of even temporary brain expertise, right, through that mental struggle, that then these AI tools can actually make that person’s thinking or learning even better down the line.
GORDON: I think it’s very plausible, as Nataliya mentioned, it’s a small sample.
But it fits in with probably six or 7 million years of human evolutionary history that we’ve been trying to find tools to help us for a long time. And sometimes you have to try the task first before you realize you need a tool or what the best tool is there.
There’s a big disconnect between human brain capacity and what we try to do with it. We have a limited immediate memory, a limited long-term memory, and the AI tools, at least in a couple of different dimensions, provide a factor of a billion better performance. So I’m not surprised that people both use them and once they have experienced the difficulty of thinking, want to use them even more.
I’ll stop there, but.
CHAKRABARTI: No, it’s okay. But so also, again, this question of what is thinking is like so large that I’m also wondering about again, going back to just writing through whatever means it might be. When we thought, when we talk about, I’m thinking of students, like when we talk about learning, what we also want is like retention of the ideas ideally, or concepts that they engage through writing.
And Nataliya, let me ask you it, with that retention, did you find that the group that had not Google and not ChatGPT, did they have the longest sort of retention of information?
KOS-MYNA: Yes. So what we actually found here, retention suffered dramatically for ChatGPT group, right?
Just to give you one number as an example, 83% of participants in ChatGPT group couldn’t quote anything 60 seconds after they gave in the essay. So not even a day. Not even a week, but 60 seconds. Compared to search engine and brain only group. But what is more important is back to the point of overall engagement.
What also, Dr. Gordon actually alluded to brain only group, and we actually don’t report this in the paper, but they provided us with so many details, verbatim details about their essays. We even didn’t ask for those, but they almost all quoted essays fully, because they were so engaged in the topics.
They wanted to share their opinions just based on one question, like, why did you pick up this topic? People do want to share. That was definitely not the case with the ChatGPT group. So that I think is very important. What is also a bit striking, and I think this is where maybe some of more doom and gloom creeps in.
Is that 15% of participants in ChatGPT group actually let us know that they do not feel any ownership towards their essays. And you would say no. Not surprising. They didn’t really write them per se, but I think this where this disconnect might be a bit dangerous and where we would want to also pay additional attention.
CHAKRABARTI: Okay. So now we get back to this idea, this truth that these large language models are here, they’re here to stay. They’re already ubiquitously available on planet Earth. And Barry Gordon, you said this earlier, like any tool, the question is how can we use it? How do we use this tool to its best?
Not its best ability, but to bring the best outcome for whatever it is that we’re trying to do. Now with that idea not that long ago. We spoke with Ethan Mollick. He’s at the Wharton School, and he’s a leading thinker about AI. And we talked actually in detail about how he uses ChatGPT or Claude or other large language models to actually write, and here’s what he said.
When I write things, I always do my own draft, but then I absolutely extensively use AI afterwards. I use it to help me do research. Some, the deep research mode of AI is quite good. Other modes can make things up. But then once I’ve done writing something, I’ll often ask the AI to act like an editor and give feedback.
I’ll ask for it to help me. … I didn’t land the last sentence, give me 30 versions of that sentence. So I can start to come up with different ideas based on that. Pretend to be a naive reader and tell me what’s confusing. So I managed to both use my own work, but also draw on the AI when I need to.
CHAKRABARTI: Barry Gordon, looking ahead, or even right now since people are using these tools, Mollick seems to be describing something that both of you have talked about. Make the brain work first. And then use the AI tools in a way to enhance the written product. What do you think about that?
GORDON: I think that’s a good strategy. You have to, there’s another part we haven’t really talked about, actually. Brains want to be lazy. Having a brain and using a brain is a very energy intensive process. And our brains have been built in part to try to take the easy way out. But we know from history of exercise you’ve got to put some effort into it to get some real, lasting benefits.
So I think his strategy of writing is the right one, and in fact, it’s the one I’ve been trying to use too. There’s a first draft of everything and then I use the AI tools to help me research it fuller. More fully.
CHAKRABARTI: But brains want to be lazy. That makes me think that it’s a real uphill battle, especially for young people who are going to inherit the world that we live.
That we’re leaving them. Like they’re more likely, not even them, anyone’s just more likely to just put the prompt in as Nataliya had talked about. We have to just train ourselves to not to do that. Barry?
GORDON: Yeah, absolutely correct. Dr. Lieberman wrote a book on Why we Exercise, which mentions that exercise was something we had to do in the past in order to just get food and survive.
Now that we don’t have to do it, we don’t really like doing it most of the time, but we have to do it now for our health. In some ways, the AI tools are going to be taking away some of the motivation, but we have to provide that ourselves.
CHAKRABARTI: Nataliya, I’d love to hear you on this.
KOS-MYNA: Yeah, absolutely. So I agree, in part absolutely with Audrey and of course like Dr. Gordon, what they are saying, and a lot of other people actually shared the same thing, Hey, I am first draft, and I ask AI to do the fix for me. But I want to just caution very carefully against one little but important note. You, me, Dr. Gordon. Dr. Mollick. We have all been born before AI.
Long before AI, right? So LLMs most, let’s say LLMs, right? Because AI was there, definitely, but LLMs, and we learned how to write, how to ask questions, how to ask difficult questions, how to structure. Or Meghna, you are in the whole workspace of fact checking, delivering news. Structuring and doing this all the time, every day.
So we already know how to do that. We are able to create those drafts. We are able to ask questions, right? I’m more concerned and I think we should all be concerned, ask important questions about is … younger adults, teenagers, kids, this new upcoming generation, are they all really fully set on learning how to extract information, work through information, create a draft, ask this question, right?
We should all be concerned, ask important questions about is … younger adults, teenagers, kids, this new upcoming generation, are they all really fully set on learning how to extract information, work through information?
The art of asking the question is as important, so I would just be cautious about just slapping AI then as a must, because that’s almost how it sounds. And let’s ensure that we have the skill. We learn the skill. We might not be the experts in this skill, like yourself, for example. But it is very important.
To keep this little detail really on the back of our minds.
CHAKRABARTI: That’s actually quite a big detail, Nataliya, and I’m glad you brought it up.
Because basically what you’re saying is there’s a whole new skillset that needs to be developed on effective use of these large language models in order to continue what both you, and Nataliya, you and Barry have been talking about. Like to get that whole brain activity going, to produce the best thinking, the least homogeneous thinking.
By the way, that part still sticks with me. And that maybe we went through the same evolution before, when we moved from an oral tradition to a written tradition. And so that’s something that I guess we have to look to schools and teachers and families in order to try and encourage that in young people as we go forward, Nataliya.
KOS-MYNA: Absolutely. I just also want to use this opportunity to give a shout out to 3,600 teachers all around the world who reached out to us after the study went up and became online and accessible and available. So many emails, and they shared so many concerns and very valid points on how these technologies should be integrated in their workspace so they can actually do their job well. And they do care. So I think it’s very important to listen to those people. They are actually truly the experts, at least if you take example of the ad tech and AI.
CHAKRABARTI: Yeah. Barry Gordon, we’ve got about 30 seconds left.
I’m going to give you the last word today.
GORDON: Oh wow. Use generative AI, carefully use it, but realize its problems, as well as its potentials. And don’t take the lazy way out and let it do everything for you. That’s all I basically have to say.
The first draft of this transcript was created by Descript, an AI transcription tool. An On Point producer then thoroughly reviewed, corrected, and reformatted the transcript before publication. The use of this AI tool creates the capacity to provide these transcripts.