TL;DR: LLM use reduces friction, leading to more superficial engagement with material, leading to more shallow argumentations.
This study explores the cognitive load and learning outcomes associated with using large language models (LLMs) versus traditional search engines for information gathering during learning. A total of 91 university students were randomly assigned to either use ChatGPT3.5 or Google to research the socio-scientific issue of nanoparticles in sunscreen to derive valid recommendations and justifications. The study aimed to investigate potential differences in cognitive load, as well as the quality and homogeneity of the students' recommendations and justifications. Results indicated that students using LLMs experienced significantly lower cognitive load. [following is the key point} However, despite this reduction, these students [that used LLMs] demonstrated lower-quality reasoning and argumentation in their final recommendations compared to those who used traditional search engines.
Carsten, this is great, thank you so much for sharing it. I will plan to add this to a future post. And I completely agree, the brain imaging is conceptually interesting, but really a potential indirect measure of things we can measure more directly. Great comment, and I appreciate you sharing a relevant paper.
Interesting. My son (7th grade) uses ChatGPT to figure out coding for physical computing....and find answers on trying to modify a dirt bike. My husband uses ChatGPT to write his emails so they sound very polished. He's an eloquent talker, but not writer. I'm the opposite (better writer than speaker) and hardly ever use ChatGPT - and wonder if I should use it after a first essay draft? I worry that it will change my "voice" in writing.
Will Storr posted something about Substack essays and ChatGPT and how people use it to post their essays - people generally like the ChatGPT essays more but the essays all start sounding the same.
Fascinating comment on all fronts. I feel like with physical computing and dirt bike modification, it's probably likely that your son is engaging in a lot of brain-first activity, where he's defining a problem and then getting some tutoring basically? I'm not sure at all, but what do you think? ...I think your concern about voice is quite valid from everything we know so far about the diversity-scrunching effect of LLMs on writing. Of course, you have the ultimate call, so you could run your writing through it and just see if it pushes you in interesting ways, particularly since you've done the brain-work first. ...I'm going to look up the Will Storr post!
So many interesting threads here. Years ago when I was directing films I worried that I might lose my voice on set with so many people offering their ideas. My wife just laughed at me. Ultimately, I came to realize that it's much easier to reject ideas I don't like than use ideas that never occur to me. Now, I kind of feel like the LLMs are similar. When using them as a glorified thesaurus, which is my main use, I almost always have pretty immediate strong reactions to suggestions, one way or another. What I do like though, is that when looking for other ways to say something, I can tailor the search much better than I ever could with a thesaurus (suggest the mood, context, what-have-you for the suggestions that the LLMs generate)
Great analogy, Rufus, thanks for sharing that. Love this line: "Ultimately, I came to realize that it's much easier to reject ideas I don't like than use ideas that never occur to me."
Regarding my son - sometimes I recommend him to just try and learn that coding language so he can tweak the ChatGPT coding to meet his needs. And he huffs.
I'm of the brain rot category and not much will change my mind at this later stage of my life. I avoid AI when I can, feeling that if you don't use your brain, you loose it. It's not about doing things better for me; it's about keeping the muscle working as much as I can and figuring out how to get places, how to write, how to do things, how to strategize and plan, how to communicate, how to retrieve information and remember things with my own brain. This is one area where I've individualized my response to new technology based on a sample size of one. That said, I do look things up for meaning and spelling etc. And I did read your piece with curiosity and interest. Well done on a complex topic which I'd prefer keeping simple for myself.
I like the framing of "time under tension" that Cleo Abram applied to our relationship with AI in her interview with Sam Altman. It's a weightlifting concept — i.e. squatting for 30 seconds is more beneficial than squatting for 3 seconds. Rather than framing what I ask AI to do as "too easy" or "not too easy", I can ask myself, is this a muscle I want to develop or not.
Great post. There are tons of great ideas in Range that make it one of my favorite books, but the learning approaches section was a highlight (I taught college at the time and could relate pretty well to the topic). Without thinking about a topic, it makes sense that you are just "copying" rather than activating the brain to learn. By thinking first, you are aware of what you don't know and then can pick that up (and are more likely to retain it) for a longer period.
Thanks for this great comment, Kevin. And I completely agree. Sometimes I try to write a draft when I know I don't have all the info, because the points where I get stuck show me what I need to go learn.
AI becomes far more powerful when it helps us clarify our thinking rather than shortcut it, especially when we use it to surface patterns we might miss on our own. I’ve been exploring that intersection in my own work, and this piece resonates strongly with that approach. Thank you ✌️
Thank you for breaking down the research about AI and how it could be affecting users. I must say that it resonates with my experiences. AI is incredibly good if you want to be more productive and up your game. I've had to adhere to my personal policy for using it to avoid making it a crutch.
Thanks Tammy - this is really useful and interesting. AI is here and it is an amazing tool, so it's really important we learn how to best use it to enhance our own thinking and work. I love using it to polish what I have already done, and to speed up the process of creating and adding content to something I have already designed.
This was fascinating. I couldn’t tell you the last time I used ChatGPT, and the times I have were for experimenting to see what it was all about. It’s not a tool that I ever think about using! But I also don’t have a day job, so needing it isn’t ingrained. Though I do know folks who use it all the time, and I do see the value in it as a tool. I think what’s most interesting about this article, is that the students weren’t actually learning when that’s all they used. That’s super interesting. And I think the point of doing. anything is to learn! To expand our horizons, and if it isn’t doing that, then it isn’t working. I wonder how the students reacted to the results? Did it change their perspective of how they will use AI in the future?
That's a great question, and I have no idea if or how it changed the students' perspective, although it certainly changed mine! I do find myself reminding myself over and over not to reflexively use it first for things that aren't akin to Google search. The domain where I find it unequivocally wonderful is tech support for my Mac.
Great content. To me, this really points out just how important the prompting process is when engaging with LLMs. If we don’t clearly think through the objective and desired outcome, how can we possibly use AI in a valuable way?
Interesting. For my own part, I tend to use Claude and Gemini, but most notably I feel that the use of any LLM is like wielding a tool. If I gave you an excavator and you had never operated one, you might well do worse damage than if I gave you a shovel. That said, once you've had a little experience, excavators are much more efficient at digging trenches.
It's taken me a while, and I feel as though I'm still only picking the low hanging fruits for now, but I believe my own writing has been dramatically improved by using Claude, even if my use could reasonably be (perhaps disparagingly) characterized as little more than a glorified Thesaurus.
I guess my question is: it would be interesting to see how the same experiments stacked up when comparing side by side, regular users of LLMs with those who don't.
That would absolutely be interesting. And this is just one study (albeit I think one that is well done, and that fits conceptually with other work). Unlike with the excavators, though, I could see arguments both ways—that regular users could be better or worse off. I'd think of it more like GPS than an excavator, you can become so reliant on it that you don't know your way around your own neighborhood. Or you can use it but also pay attention. I can't entirely convince myself either way, so would love to see more work of this nature.
Interesting analogy, of course, even with new skills pushing old ones out, it is possible that the old ones just become obsolete (very few of us can ride horses these days, but transport is a lot more efficient). I guess there is the risk of falling prey to what we can measure and missing what we can't: we can grade an essay, but I don't think it's so clear what other skills the new tool unlocks yet (again, in my own limited use, I'm pretty sure that, even on the old writing metrics, with some fluency with the tool I'm producing better output. Perhaps more significantly, I'm producing it much faster which frees my time to learn higher leverage skills.
To be clear, I wouldn't want to be bucketed as an evangelist, but I do think it's easier to cut new technologies down rather than identifying how they might change the world. To your GPS example, we might all be a little worse at personal navigation without our phones, but with them we now have ride-sharing and any number of magical possibilities thanks to GPS tech.
Late to the party here, but on the shovel vs excavator analogy, it is pretty clear that it is desirable to use the excavator in the name of productivity if you're trying to get as many trenches dug as possible. But the human body is built to dig trenches/ generally be active. If we only ever use the excavator, we're going to be pretty unhealthy and probably unhappy because we're meant to move.
I suspect there may be something similar with respect to offloading cognitive tasks. I own a book of short works by Richard Feynman titled The Pleasure of Figuring Things Out, which is the generally notion I'm pointing to here. So if we let our "figuring things out" muscles atrophy, I'd bet that will have all sorts of negative effects, similar the physical atrophy of not digging trenches anymore.
That said, I'm an avid cyclist and I certainly appreciate that all the productivity gains in my life/ across society have afforded me the opportunity to be physically active in that form, rather than digging trenches. So maybe there is a world where, for example, offloading the composition of an e-mail to an LLM creates space for me to engage in some other more creative pursuit. But then again, I think we're generally wired to avoid hardship/ challenges, and only about a quarter of people manage to meet even the minimum standards for routine physical exercise.
Hey Dan! great notes. I have a couple of contextualizing thoughts around excavators as they have become nearer and dearer to my own heart, quite by accident (a perhaps inopportune turn of phrase to use when talking about heavy machinery, but it's done). Firstly, they are more work physically to use than one might expect, and secondly, there is joy in using them. I say this as an amateur user, who had no idea that their use was on my own bucket list until after I inadvertently had a go at using one. It happened about a year ago, when my electrician offered me the opportunity to use his as he watched me toiling in a trench. My wife now refers to my efforts moving big rocks about as the latte circuit (meaning not requiring the real effort that moving rocks normally requires--actually, the rocks I'm moving now I could barely inch with a crow-bar), but I can assure you that I'm constantly hopping up and down from the cab, and at the end of a day I'm exhausted.
As for a physical analogue of Feynman's book, a few years ago I was putting up a fence out in the bush (I confess I'd hired someone else to set the posts), a not insignificant task as we have about a quarter mile along the road. Anyway, I found myself out there one day, sawing and drilling, and, for a moment, I wondered why I hadn't just hired someone for the whole task ... then it dawned on me that I was happy doing the physical work. Being out in the great outdoors, surrounded by wildlife. It was a bit of an epiphany for me to realize how much I enjoy hard physical work.
On the cycling front, I too love cycling, and one modern addition to that mode of transport that I'm really happy about is the advent of electric bikes. I don't actually have one myself, but I love the knock-on effect of others adopting them: a proliferation of bike lanes at home (Santa Monica) and general biking culture that I've long envied in other other countries.
Great post! I think SO much about this with my students. For my 7th graders I've fallen pretty strongly in the, "You shouldn't use this at all in my class" camp. Lately, though, I've been thinking I have some responsibility to teach them how to use this tool responsibly because they'll be encountering it in the outside world whether in my class or not (and, after sitting in on some Sex Ed classes recently, my current thinking is starting to sound akin to an abstinence-only talk haha). I guess my question is this: I know your son is quite a bit younger, but at what age would you want his teachers to start introducing ChatGPT and other AI-infused tools to him? Do you think your answer is representative of what most parents what say?
Also thanks for sharing the link to the Q&A at the end. I missed it but will be sure to read it!
Hey Matt, first off, I just added a footnote to this post because another teacher made interesting comments. Here's the footnote in case you have thoughts:
"Justin Cerenzia made a thoughtful comment on this paragraph, so I want to share it here: 'We’re seeing similar things as we pilot an LLM-assisted (question generation based on student writing) oral assessment platform (with a video component). But I might quibble with this part [that for students the point is the act of writing, not making the best finished product]. Especially in secondary schools, the point is the product. So they become competent at producing a product, but not, necessarily, at learning.'”
Beyond that, what I'm hearing from you is that abstinence-only education works every time. Amirite?? ...At what age would I want my teachers to start introducing AI tools. Gosh, tough one. I mean, I wouldn't be averse currently to a tool that allowed my kid to verbally ask questions about a topic he's interested in. Because that's what he does to me, and I'm not sure I'm that much different from ChatGPT in that regard. Well, that's not totally true, at this age I'm not just giving him information, I'm explaining some context, linking it to other things he understands, and getting him to engage in collective looking up of stuff. (Basically overnight he got on a reference book kick and now has a stack almost his height—granted started with an enormous dictionary I've had for years. He had me reading the index of an atlas in order recently..."Aruba, Antigua," etc. etc.) But I think I'm probably similar to a chatbot in many respects when it comes to just answering informational queries. I could also see some great uses in tutoring for math or language. But it's different when the idea is producing something or grappling with something and it's just a shortcut. Frankly, I don't see a ton of downside in a big delay on that front. It isn't hard to use a chatbot, and I think it will only get easier, so this doesn't feel like the kind of thing to me that kids must learn how to use now, whereas I think their brains are things they must learn how to use now. But given that abstinence-only won't work, maybe starting in middle school I'd have a class, maybe with something like the essay-based work above, where you get kids used to doing their own essay and then using AI after that. Perhaps start acclimating them to that sequence, and then expanding it a bit every year. I don't know, what do you think? And honestly I have no clue what other parents would say, but now I'm going to ask a few.
Oooooh, I'm not sure I agree that it's easy to use a chatbot. Superficially, yes. But, to me, that's a little analogous to saying "it's easy to ask people questions.", and I'd guess pounds to pence that, as a reporter, you don't agree with that. I think there is skill in extracting what you want, and learning that earlier is a very high leverage skill to acquire that, frankly, goes well beyond interactions with LLMs.
Anyway, just my two cents (to mix currency metaphors). Perhaps easy for me to say, since my kids are both at university now, but I would encourage kids very young to interact with chatbots. I'm not sure I can see any argument that makes sense to exclude the use any longer than you'd exclude access to say video content (I mean, at interactions with LLMs are less passive than watching TV, no ...?)
Fair points and well put. Although, I will say that when I was less experienced I was more interested in all sorts of specific interview tactics, and I actually now think that most of them don't make much difference, and doing your homework and genuine curiosity is most of the ballgame. (One way I think this manifests: If someone is genuinely curious, when something unexpected pops up, they keep asking about it even if it wasn't in the script.) And as far as video content, I do think much of that is distressingly passive. Not all of it, but I think TV in general is probably very overwhelmingly passive consumption, so for me that's a very low bar. I recall some old work showing that most people use the internet for entertainment but some people use it primarily for information. I expect there will be many devils in the details of how LLMs are used; structured tutoring for example seems to me to have a ton of promise. But given how much really young kids have to learn, I see more risk than benefit at very early ages for unfettered use. But I'm curious and don't have my mind made up! I think this is motivating to look for more research and see what is done so far, even though it's very early to have strong evidence.
Very interesting footnote. I actually think it is quite a nice feature of my job that my kids are still young enough where the point is still the act of writing rather than the best finished product. It actually makes my grading job easier where I grade more for process/structure of writing rather than how correct the argument is. But that aside...
I think that sort of verbally-ask-questions thing would be pretty good and age appropriate, but is it luddite-y of me to say I still think your reference book kick and having him look things up is far preferable? Something about the delayed gratification and friction between question and answer appeals to me. That's hilarious he had you read the index in order. I watched an episode of 30 Rock last night where Tracy reads the phone book on stage to great applause ("A triumph!" someone shouted out lol), and your reading the atlas reminds me of that haha.
Overall, I agree that there's nuance with each subject, and I like your suggestion of maybe introducing this in middle school with different drafts. I think I could pull this off, though I suspect some kids would abuse it (just yesterday, a kid handed in an essay done entirely by ChatGPT). I do think a lot of kids would be able to handle this though, and it would be a net positive to introduce it to them in this way. So I think maybe I should at least try it. I have plenty of intrinsically motivated kids who it feels like a disservice to not show them this, but I don't know. Maybe I should ask ChatGPT what it thinks!
Hey Matt, I hear your reticence, but just to play devil's advocate, here's one thing I do like about LLMs that reference books don't give you (and it's a thing that as a pedagogue I strive to include when I'm teaching anyone): the LLMs often ask follow up questions, which I think is really helpful, both to model the habit of asking follow up questions, but also to nudge thinking beyond the immediate question. I'd be curious of your thinking on that.
To play...devil's devil's advocate (?), when I'm reading reference books with my son, there is no way either of us are escaping the other's follow up questions until basically it's past bedtime;) I think probably the most useful thing I do (and not saying this is unique to me or even to human intelligence) is constantly connect things to other stuff he already knows. One could argue that I'm like a very well tailored LLM in that sense, I suppose.
(devil's advocate)^3 ... your son is one of the luckiest children in the world. While it would be wonderful if all parents engaged as you clearly do, my experience (having raised two kids and watching carefully how other parents interact) is that you are actually unusual. LLMs are almost certainly a pale substitute for you, but I suspect they are a pretty amazing alternative for the kid who's parents leave the house at 6am and return exhausted at 8pm, which happens at many different points on the socio economic spectrum. To reverse your observation, LLMs are a better-than-nothing substitute of you ;)
Btw, apologies for monopolizing this comment section. Obviously, this is something that I'm very interested in, and I think tutors are perhaps the most promising positive outcome of LLMs (maybe with medical a close second). I pitched as much a couple of years ago in the commentary section for the fourth chapter of my time-travel book The Curve of Time (check it out:
Don't apologize I love this stuff! This is a great use of the comment section. And I think I agree about tutoring. At some point I want to dive into the literature a bit more and get into more specifics about what we know so far regarding specifically when LLM use is better or worse. Even from a cursory look I can see there is plenty of nuance, as there was in the study in this post. And you're making me more motivated to work on such a posts (or posts, since I expect we'll be updating our thinking regularly). I will check out the link, sounds fascinating! But most importantly: this is an excellent conversation that is stimulating my thinking, and a great use of the comment section, so you should believe me when I say that there is no quota on thoughtful comments around here.
I appreciate this comment, Rufus! I genuinely hadn't thought about it, but it makes a lot of sense, especially with how you put it later about it being a good substitute for kids whose parents can't be like David.
Socrates predicted that writing would erode memory. The typewriter and keyboard accelerated that erosion, and now AI is threatening to finish the job.
Research (Mueller & Oppenheimer, 2014) shows that this isn't just about AI—it's about friction. They found that typists tend to mindlessly 'transcribe' lectures verbatim, while handwriters—forced by the slowness of the pen—must engage in 'desirable difficulty,' synthesizing and summarizing concepts in real-time.
To avoid the 'cognitive debt' you describe, I use a hybrid workflow: Analog for thinking, Digital for execution.
I handwrite the initial ideas to force that cognitive synthesis. Once the structure is locked, then I switch to the keyboard/AI for speed and scale. Slow down to understand; speed up to deliver.
I actually find this paper more informative - clear and rigorous, and easier to interpret (without all the brain imaging stuff)
https://www.sciencedirect.com/science/article/pii/S0747563224002541
TL;DR: LLM use reduces friction, leading to more superficial engagement with material, leading to more shallow argumentations.
This study explores the cognitive load and learning outcomes associated with using large language models (LLMs) versus traditional search engines for information gathering during learning. A total of 91 university students were randomly assigned to either use ChatGPT3.5 or Google to research the socio-scientific issue of nanoparticles in sunscreen to derive valid recommendations and justifications. The study aimed to investigate potential differences in cognitive load, as well as the quality and homogeneity of the students' recommendations and justifications. Results indicated that students using LLMs experienced significantly lower cognitive load. [following is the key point} However, despite this reduction, these students [that used LLMs] demonstrated lower-quality reasoning and argumentation in their final recommendations compared to those who used traditional search engines.
Carsten, this is great, thank you so much for sharing it. I will plan to add this to a future post. And I completely agree, the brain imaging is conceptually interesting, but really a potential indirect measure of things we can measure more directly. Great comment, and I appreciate you sharing a relevant paper.
Interesting. My son (7th grade) uses ChatGPT to figure out coding for physical computing....and find answers on trying to modify a dirt bike. My husband uses ChatGPT to write his emails so they sound very polished. He's an eloquent talker, but not writer. I'm the opposite (better writer than speaker) and hardly ever use ChatGPT - and wonder if I should use it after a first essay draft? I worry that it will change my "voice" in writing.
Will Storr posted something about Substack essays and ChatGPT and how people use it to post their essays - people generally like the ChatGPT essays more but the essays all start sounding the same.
Fascinating comment on all fronts. I feel like with physical computing and dirt bike modification, it's probably likely that your son is engaging in a lot of brain-first activity, where he's defining a problem and then getting some tutoring basically? I'm not sure at all, but what do you think? ...I think your concern about voice is quite valid from everything we know so far about the diversity-scrunching effect of LLMs on writing. Of course, you have the ultimate call, so you could run your writing through it and just see if it pushes you in interesting ways, particularly since you've done the brain-work first. ...I'm going to look up the Will Storr post!
So many interesting threads here. Years ago when I was directing films I worried that I might lose my voice on set with so many people offering their ideas. My wife just laughed at me. Ultimately, I came to realize that it's much easier to reject ideas I don't like than use ideas that never occur to me. Now, I kind of feel like the LLMs are similar. When using them as a glorified thesaurus, which is my main use, I almost always have pretty immediate strong reactions to suggestions, one way or another. What I do like though, is that when looking for other ways to say something, I can tailor the search much better than I ever could with a thesaurus (suggest the mood, context, what-have-you for the suggestions that the LLMs generate)
Great analogy, Rufus, thanks for sharing that. Love this line: "Ultimately, I came to realize that it's much easier to reject ideas I don't like than use ideas that never occur to me."
I think it was this one. https://substack.com/@willstorr/p-163648319
Regarding my son - sometimes I recommend him to just try and learn that coding language so he can tweak the ChatGPT coding to meet his needs. And he huffs.
I'm of the brain rot category and not much will change my mind at this later stage of my life. I avoid AI when I can, feeling that if you don't use your brain, you loose it. It's not about doing things better for me; it's about keeping the muscle working as much as I can and figuring out how to get places, how to write, how to do things, how to strategize and plan, how to communicate, how to retrieve information and remember things with my own brain. This is one area where I've individualized my response to new technology based on a sample size of one. That said, I do look things up for meaning and spelling etc. And I did read your piece with curiosity and interest. Well done on a complex topic which I'd prefer keeping simple for myself.
Excessive Ease vs Desirable Difficulty
Oh man, I'm so annoyed I didn't go with alliteration!
All too deep for me....but I'll offer "undesirable facility"
I like it. That has better rhythm than my version.
I like the framing of "time under tension" that Cleo Abram applied to our relationship with AI in her interview with Sam Altman. It's a weightlifting concept — i.e. squatting for 30 seconds is more beneficial than squatting for 3 seconds. Rather than framing what I ask AI to do as "too easy" or "not too easy", I can ask myself, is this a muscle I want to develop or not.
Interesting analogy, and I love a good analogy. Thanks for sharing this, Renita. It will definitely stick in my mind.
Great post. There are tons of great ideas in Range that make it one of my favorite books, but the learning approaches section was a highlight (I taught college at the time and could relate pretty well to the topic). Without thinking about a topic, it makes sense that you are just "copying" rather than activating the brain to learn. By thinking first, you are aware of what you don't know and then can pick that up (and are more likely to retain it) for a longer period.
Thanks for this great comment, Kevin. And I completely agree. Sometimes I try to write a draft when I know I don't have all the info, because the points where I get stuck show me what I need to go learn.
AI becomes far more powerful when it helps us clarify our thinking rather than shortcut it, especially when we use it to surface patterns we might miss on our own. I’ve been exploring that intersection in my own work, and this piece resonates strongly with that approach. Thank you ✌️
Well said, David, and thanks for reading!
Thank you for breaking down the research about AI and how it could be affecting users. I must say that it resonates with my experiences. AI is incredibly good if you want to be more productive and up your game. I've had to adhere to my personal policy for using it to avoid making it a crutch.
That's a great way to put it, and I'm in the same boat, Daniel.
Thanks Tammy - this is really useful and interesting. AI is here and it is an amazing tool, so it's really important we learn how to best use it to enhance our own thinking and work. I love using it to polish what I have already done, and to speed up the process of creating and adding content to something I have already designed.
This was fascinating. I couldn’t tell you the last time I used ChatGPT, and the times I have were for experimenting to see what it was all about. It’s not a tool that I ever think about using! But I also don’t have a day job, so needing it isn’t ingrained. Though I do know folks who use it all the time, and I do see the value in it as a tool. I think what’s most interesting about this article, is that the students weren’t actually learning when that’s all they used. That’s super interesting. And I think the point of doing. anything is to learn! To expand our horizons, and if it isn’t doing that, then it isn’t working. I wonder how the students reacted to the results? Did it change their perspective of how they will use AI in the future?
That's a great question, and I have no idea if or how it changed the students' perspective, although it certainly changed mine! I do find myself reminding myself over and over not to reflexively use it first for things that aren't akin to Google search. The domain where I find it unequivocally wonderful is tech support for my Mac.
Great content. To me, this really points out just how important the prompting process is when engaging with LLMs. If we don’t clearly think through the objective and desired outcome, how can we possibly use AI in a valuable way?
I completely agree, and hope to write more about the nuances of how our specific use tactics matter.
Interesting. For my own part, I tend to use Claude and Gemini, but most notably I feel that the use of any LLM is like wielding a tool. If I gave you an excavator and you had never operated one, you might well do worse damage than if I gave you a shovel. That said, once you've had a little experience, excavators are much more efficient at digging trenches.
It's taken me a while, and I feel as though I'm still only picking the low hanging fruits for now, but I believe my own writing has been dramatically improved by using Claude, even if my use could reasonably be (perhaps disparagingly) characterized as little more than a glorified Thesaurus.
I guess my question is: it would be interesting to see how the same experiments stacked up when comparing side by side, regular users of LLMs with those who don't.
That would absolutely be interesting. And this is just one study (albeit I think one that is well done, and that fits conceptually with other work). Unlike with the excavators, though, I could see arguments both ways—that regular users could be better or worse off. I'd think of it more like GPS than an excavator, you can become so reliant on it that you don't know your way around your own neighborhood. Or you can use it but also pay attention. I can't entirely convince myself either way, so would love to see more work of this nature.
Interesting analogy, of course, even with new skills pushing old ones out, it is possible that the old ones just become obsolete (very few of us can ride horses these days, but transport is a lot more efficient). I guess there is the risk of falling prey to what we can measure and missing what we can't: we can grade an essay, but I don't think it's so clear what other skills the new tool unlocks yet (again, in my own limited use, I'm pretty sure that, even on the old writing metrics, with some fluency with the tool I'm producing better output. Perhaps more significantly, I'm producing it much faster which frees my time to learn higher leverage skills.
To be clear, I wouldn't want to be bucketed as an evangelist, but I do think it's easier to cut new technologies down rather than identifying how they might change the world. To your GPS example, we might all be a little worse at personal navigation without our phones, but with them we now have ride-sharing and any number of magical possibilities thanks to GPS tech.
Late to the party here, but on the shovel vs excavator analogy, it is pretty clear that it is desirable to use the excavator in the name of productivity if you're trying to get as many trenches dug as possible. But the human body is built to dig trenches/ generally be active. If we only ever use the excavator, we're going to be pretty unhealthy and probably unhappy because we're meant to move.
I suspect there may be something similar with respect to offloading cognitive tasks. I own a book of short works by Richard Feynman titled The Pleasure of Figuring Things Out, which is the generally notion I'm pointing to here. So if we let our "figuring things out" muscles atrophy, I'd bet that will have all sorts of negative effects, similar the physical atrophy of not digging trenches anymore.
That said, I'm an avid cyclist and I certainly appreciate that all the productivity gains in my life/ across society have afforded me the opportunity to be physically active in that form, rather than digging trenches. So maybe there is a world where, for example, offloading the composition of an e-mail to an LLM creates space for me to engage in some other more creative pursuit. But then again, I think we're generally wired to avoid hardship/ challenges, and only about a quarter of people manage to meet even the minimum standards for routine physical exercise.
Hey Dan! great notes. I have a couple of contextualizing thoughts around excavators as they have become nearer and dearer to my own heart, quite by accident (a perhaps inopportune turn of phrase to use when talking about heavy machinery, but it's done). Firstly, they are more work physically to use than one might expect, and secondly, there is joy in using them. I say this as an amateur user, who had no idea that their use was on my own bucket list until after I inadvertently had a go at using one. It happened about a year ago, when my electrician offered me the opportunity to use his as he watched me toiling in a trench. My wife now refers to my efforts moving big rocks about as the latte circuit (meaning not requiring the real effort that moving rocks normally requires--actually, the rocks I'm moving now I could barely inch with a crow-bar), but I can assure you that I'm constantly hopping up and down from the cab, and at the end of a day I'm exhausted.
As for a physical analogue of Feynman's book, a few years ago I was putting up a fence out in the bush (I confess I'd hired someone else to set the posts), a not insignificant task as we have about a quarter mile along the road. Anyway, I found myself out there one day, sawing and drilling, and, for a moment, I wondered why I hadn't just hired someone for the whole task ... then it dawned on me that I was happy doing the physical work. Being out in the great outdoors, surrounded by wildlife. It was a bit of an epiphany for me to realize how much I enjoy hard physical work.
On the cycling front, I too love cycling, and one modern addition to that mode of transport that I'm really happy about is the advent of electric bikes. I don't actually have one myself, but I love the knock-on effect of others adopting them: a proliferation of bike lanes at home (Santa Monica) and general biking culture that I've long envied in other other countries.
I enjoyed reading your article. Thanks
Thanks Paul!
Great post! I think SO much about this with my students. For my 7th graders I've fallen pretty strongly in the, "You shouldn't use this at all in my class" camp. Lately, though, I've been thinking I have some responsibility to teach them how to use this tool responsibly because they'll be encountering it in the outside world whether in my class or not (and, after sitting in on some Sex Ed classes recently, my current thinking is starting to sound akin to an abstinence-only talk haha). I guess my question is this: I know your son is quite a bit younger, but at what age would you want his teachers to start introducing ChatGPT and other AI-infused tools to him? Do you think your answer is representative of what most parents what say?
Also thanks for sharing the link to the Q&A at the end. I missed it but will be sure to read it!
Hey Matt, first off, I just added a footnote to this post because another teacher made interesting comments. Here's the footnote in case you have thoughts:
"Justin Cerenzia made a thoughtful comment on this paragraph, so I want to share it here: 'We’re seeing similar things as we pilot an LLM-assisted (question generation based on student writing) oral assessment platform (with a video component). But I might quibble with this part [that for students the point is the act of writing, not making the best finished product]. Especially in secondary schools, the point is the product. So they become competent at producing a product, but not, necessarily, at learning.'”
Beyond that, what I'm hearing from you is that abstinence-only education works every time. Amirite?? ...At what age would I want my teachers to start introducing AI tools. Gosh, tough one. I mean, I wouldn't be averse currently to a tool that allowed my kid to verbally ask questions about a topic he's interested in. Because that's what he does to me, and I'm not sure I'm that much different from ChatGPT in that regard. Well, that's not totally true, at this age I'm not just giving him information, I'm explaining some context, linking it to other things he understands, and getting him to engage in collective looking up of stuff. (Basically overnight he got on a reference book kick and now has a stack almost his height—granted started with an enormous dictionary I've had for years. He had me reading the index of an atlas in order recently..."Aruba, Antigua," etc. etc.) But I think I'm probably similar to a chatbot in many respects when it comes to just answering informational queries. I could also see some great uses in tutoring for math or language. But it's different when the idea is producing something or grappling with something and it's just a shortcut. Frankly, I don't see a ton of downside in a big delay on that front. It isn't hard to use a chatbot, and I think it will only get easier, so this doesn't feel like the kind of thing to me that kids must learn how to use now, whereas I think their brains are things they must learn how to use now. But given that abstinence-only won't work, maybe starting in middle school I'd have a class, maybe with something like the essay-based work above, where you get kids used to doing their own essay and then using AI after that. Perhaps start acclimating them to that sequence, and then expanding it a bit every year. I don't know, what do you think? And honestly I have no clue what other parents would say, but now I'm going to ask a few.
Oooooh, I'm not sure I agree that it's easy to use a chatbot. Superficially, yes. But, to me, that's a little analogous to saying "it's easy to ask people questions.", and I'd guess pounds to pence that, as a reporter, you don't agree with that. I think there is skill in extracting what you want, and learning that earlier is a very high leverage skill to acquire that, frankly, goes well beyond interactions with LLMs.
Anyway, just my two cents (to mix currency metaphors). Perhaps easy for me to say, since my kids are both at university now, but I would encourage kids very young to interact with chatbots. I'm not sure I can see any argument that makes sense to exclude the use any longer than you'd exclude access to say video content (I mean, at interactions with LLMs are less passive than watching TV, no ...?)
Fair points and well put. Although, I will say that when I was less experienced I was more interested in all sorts of specific interview tactics, and I actually now think that most of them don't make much difference, and doing your homework and genuine curiosity is most of the ballgame. (One way I think this manifests: If someone is genuinely curious, when something unexpected pops up, they keep asking about it even if it wasn't in the script.) And as far as video content, I do think much of that is distressingly passive. Not all of it, but I think TV in general is probably very overwhelmingly passive consumption, so for me that's a very low bar. I recall some old work showing that most people use the internet for entertainment but some people use it primarily for information. I expect there will be many devils in the details of how LLMs are used; structured tutoring for example seems to me to have a ton of promise. But given how much really young kids have to learn, I see more risk than benefit at very early ages for unfettered use. But I'm curious and don't have my mind made up! I think this is motivating to look for more research and see what is done so far, even though it's very early to have strong evidence.
Very interesting footnote. I actually think it is quite a nice feature of my job that my kids are still young enough where the point is still the act of writing rather than the best finished product. It actually makes my grading job easier where I grade more for process/structure of writing rather than how correct the argument is. But that aside...
I think that sort of verbally-ask-questions thing would be pretty good and age appropriate, but is it luddite-y of me to say I still think your reference book kick and having him look things up is far preferable? Something about the delayed gratification and friction between question and answer appeals to me. That's hilarious he had you read the index in order. I watched an episode of 30 Rock last night where Tracy reads the phone book on stage to great applause ("A triumph!" someone shouted out lol), and your reading the atlas reminds me of that haha.
Overall, I agree that there's nuance with each subject, and I like your suggestion of maybe introducing this in middle school with different drafts. I think I could pull this off, though I suspect some kids would abuse it (just yesterday, a kid handed in an essay done entirely by ChatGPT). I do think a lot of kids would be able to handle this though, and it would be a net positive to introduce it to them in this way. So I think maybe I should at least try it. I have plenty of intrinsically motivated kids who it feels like a disservice to not show them this, but I don't know. Maybe I should ask ChatGPT what it thinks!
Hey Matt, I hear your reticence, but just to play devil's advocate, here's one thing I do like about LLMs that reference books don't give you (and it's a thing that as a pedagogue I strive to include when I'm teaching anyone): the LLMs often ask follow up questions, which I think is really helpful, both to model the habit of asking follow up questions, but also to nudge thinking beyond the immediate question. I'd be curious of your thinking on that.
To play...devil's devil's advocate (?), when I'm reading reference books with my son, there is no way either of us are escaping the other's follow up questions until basically it's past bedtime;) I think probably the most useful thing I do (and not saying this is unique to me or even to human intelligence) is constantly connect things to other stuff he already knows. One could argue that I'm like a very well tailored LLM in that sense, I suppose.
(devil's advocate)^3 ... your son is one of the luckiest children in the world. While it would be wonderful if all parents engaged as you clearly do, my experience (having raised two kids and watching carefully how other parents interact) is that you are actually unusual. LLMs are almost certainly a pale substitute for you, but I suspect they are a pretty amazing alternative for the kid who's parents leave the house at 6am and return exhausted at 8pm, which happens at many different points on the socio economic spectrum. To reverse your observation, LLMs are a better-than-nothing substitute of you ;)
Btw, apologies for monopolizing this comment section. Obviously, this is something that I'm very interested in, and I think tutors are perhaps the most promising positive outcome of LLMs (maybe with medical a close second). I pitched as much a couple of years ago in the commentary section for the fourth chapter of my time-travel book The Curve of Time (check it out:
https://www.writtenbyrufus.com/episodes/cot-chapter-14
Thanks as always for your thoughtful curation of ideas!
Don't apologize I love this stuff! This is a great use of the comment section. And I think I agree about tutoring. At some point I want to dive into the literature a bit more and get into more specifics about what we know so far regarding specifically when LLM use is better or worse. Even from a cursory look I can see there is plenty of nuance, as there was in the study in this post. And you're making me more motivated to work on such a posts (or posts, since I expect we'll be updating our thinking regularly). I will check out the link, sounds fascinating! But most importantly: this is an excellent conversation that is stimulating my thinking, and a great use of the comment section, so you should believe me when I say that there is no quota on thoughtful comments around here.
I appreciate this comment, Rufus! I genuinely hadn't thought about it, but it makes a lot of sense, especially with how you put it later about it being a good substitute for kids whose parents can't be like David.
It's no longer if you AI, it's when.
So fun watching and participating in these amazing times 😎👍
Socrates predicted that writing would erode memory. The typewriter and keyboard accelerated that erosion, and now AI is threatening to finish the job.
Research (Mueller & Oppenheimer, 2014) shows that this isn't just about AI—it's about friction. They found that typists tend to mindlessly 'transcribe' lectures verbatim, while handwriters—forced by the slowness of the pen—must engage in 'desirable difficulty,' synthesizing and summarizing concepts in real-time.
To avoid the 'cognitive debt' you describe, I use a hybrid workflow: Analog for thinking, Digital for execution.
I handwrite the initial ideas to force that cognitive synthesis. Once the structure is locked, then I switch to the keyboard/AI for speed and scale. Slow down to understand; speed up to deliver.