Plain English with Derek Thompson - What 400,000 Essays Reveal About AI and Creativity

Episode Date: June 5, 2026

For generations, we've defined creativity by its products: the novel, the painting, the song, the breakthrough idea. We look at the work, and from the work we see the creator as “creative.” But AI is getting remarkably good at producing creative work. In some cases, experts now prefer AI-generated writing to work created by humans and can't reliably tell the difference between the two. In fact, a major literary prize even recently honored a work that was largely written by AI. It all raises a deeper question than whether or not AI can write well. It forces us to reconsider what creativity actually is. Today, neuroscientist Adam Green joins the show to discuss how AI is changing the way we write, think, and generate ideas. His research finds that while AI can make our language more polished and sophisticated, it may also make our thinking more uniform. The sentences get sharper. The ideas get more predictable. And If creativity is no longer something we can recognize from the final product alone, we may need a new, more human definition. Subscribe to our YouTube channel here:https://www.youtube.com/@PlainEnglishwithDerekThompson If you have questions, observations, or ideas for future episodes, email us at [e-mail]. Host: Derek ThompsonGuest: Dr. Adam GreenProducer: Devon BaroldiAdditional Production Support: Ben Glicksman Learn more about your ad choices. Visit podcastchoices.com/adchoices

Two weeks ago, the Commonwealth Short Story Prize, one of the most prestigious awards in literary fiction, announced its annual winners. There was just one problem with this year's list of winners. One winning story, it turned out, was largely written by artificial intelligence. It's tempting to think that the prize judges in this case must have just been stupid, that AI sucks at writing just like it sucks at everything else in the creative realm, and surely any expert in writing worth their salt would have noticed the difference. But the evidence suggests otherwise. In a 2025 study by researchers at Stony Brook, Columbia, and the University of Michigan,

three leading AI models were pitted against MFA-trained writers. In initial tests, expert readers clearly preferred the human writing. Sounds reassuring. But once researchers fine-tuned chat GPT on an individual author's full body of work, the results flipped. Experts suddenly preferred the AI's writing and often couldn't tell whether it came from a human or machine. This raises a deep question, I think, that cuts to the heart of what we think creativity actually is.

We've always defined creativity by pointing at artistic products, paintings, books, plays. We decide that Ulysses is creative, or Hamilton is creative, or Gernica is creative, and then we reason backward to call their creators creative, James Joyce, Lynn Manuel Miranda, Picasso. So, creativity in this framework, is inferred by the product. But what happens when experts decide that a short story is creative, and it turns out that the author of that short story is a chips rack in a data center in Virginia? As Rebecca Winthrop wrote recently in a New York Times essay, quote, for the first time in human history, we have a technology that can generate words separately from the thoughts they

represent." And that means I think we have to reconceptualize what creativity means. We have to think of it as not just a product that could come from any source, carbon-based or silicon, but as a process that must be indelibly

human. For the last eight years, today's guest, Georgetown University's neuroscientist Adam Green, has been leading a national research team tracking the range of novel ideas that students put into their college application. essays. In one study, he and his team examined personal statements from hundreds of thousands of

student applications before and after chat GPT arrived in November 2022. What he found was striking. After chat chitb became available, the essays used richer and more diverse language. External experts even rated them as more creative. But something was missing. Something that turns out to be entirely central to what creativity might actually be. The essays became more similar to each other. They explored a narrower range of ideas. Students who use AI, Green argues, are making a Faustian bargain. Their prose gets refined, but their ideas get basic. Their sentences, in some cases, get sharper, but their minds get duller. And as more of us lean on these tools, not just students, but writers, workers, anyone who thinks for a living, we all incur that same risk.

Sharp sentences for duller minds. I'm Derek Thompson. This is plain English. Adam Green, welcome with the show. Thanks, Derek. Great to be with you. Tell me who you are and tell me what is the question that you're motivated by as a researcher? What are you spending your career trying to figure out? I'm a neuroscientist.

So I study how brains do a lot of the interesting things that they do. To me, the most interesting thing that brains do is come up with creative ideas. So I've spent a lot of time mapping out how creative systems work in brains, and then a lot of time figuring out how we can help those systems work better by things that we do in classrooms, things that we do in labs where we actually zap brains with various forms of stimulation to help those systems work better. And then recently, there's this new kind of creativity.

Artificial intelligence is generating ideas that we're calling creative. And I think that there's a good case to be made for that. But one of the things that's been fascinating to me in my lab is how that's happening differently in these artificial systems than it happens in our organic systems. And what that might mean for the nature of creativity. So that's something that's really been motivating. along with still zapping a lot of brains. Before we get into the paper that first drew my attention for this episode

on student admission essays and artificial intelligence, what you just said sort of bump me in a way, and I love you to spell it out a little bit more. What is creativity and what do we know about how it is produced organically in the brain? So the question of what is creativity is a big one, and an old one. And somewhat hubristically,

a friend of mine, James Kaufman, who's another creativity researcher, he and I have been leading 200 creativity researchers from around the world, people who have made the terrible life choice to think about this stuff all the time. In trying to answer this question,

so it's a project that we call the Creativity Ontology Project, and we're trying to map out exactly what creativity means in a way that not just researchers, but teachers and people in industry can understand it. I think I will let you know when we get to the end of that journey, just what the creative world or the world of creativity research thinks creativity is.

But what I think matters about your question is what kind of creativity you're looking for. And so I think the definition of creativity has tended to slid, plant toward the product. And what I mean by that is we think of an invention, right, or a song. And we say, well, that's really creative. And we have a pretty well-agreed-upon standard for what we mean by creative when we're talking about a product like an invention or a song. And that's two criteria, novel and useful, right?

What's interesting to me, and this comes back to the AI question, is those are fine descriptors for a product. They're not very good descriptors if what we mean by creativity is the process of creating. Creativity, the process of generating something. And I look at brains, right? And what brains are doing, right?

Those descriptors, novel or useful, have nothing to do with the processes happening in your brain, right? You can't use it to open a can, right? And it's not novel. in the sense that these same systems do a lot of things, right? And they've done all of those things 50 times today before you were writing your song, right? So neither novel nor useful is a particularly good description

of creativity as a process. So we've done some work looking at what we mean by creativity as a process, and it has a lot to do with our search through our own semantic networks, our meanings, right? And how that happens in a way that, focuses our attention inward, and that is constrained by a generative goal. So that's, if you're asking me what I mean by creativity, and I'm right now, as, as you know, at the annual meeting of the Society for the Neuroscience of Creativity, and there are a bunch of people

here who love to disagree about what creativity means. But if you're asking me, and I think more and more, as we've moved towards studying creativity and brains, the importance of understanding creativity as a process is being brought more to the fore. I think those are the elements that matter the most. I think this distinction between creative output and creative work is going to be one that we return to a lot because I know it's a theme of your research on artificial intelligence. I think even just to skip a little bit ahead in our conversation, artificial intelligence often allows people who are not doing creative work to nonetheless produce pieces of writing,

even music, that might sound to an external observer as creative. And that creates this like existential question of, well, if there was no creative work that was done, is the output itself an example of creativity? Like it's a great and rich, like existential question, definitional question. Even I, to use the, you know, $10 word that you use to name your, you're the name of your research, an ontological question. What is creativity if we can produce creativity without doing creative work? Let's jump right into it.

Let's talk about your study, which is, for those who want to look it up at home, quote, the creative link between words and ideas is weakening in the AI era. And here you looked at hundreds of thousands of college admissions essays before and after the introduction of chat chv-t. Tell me what you found. What did these essays gain in the post-chat chavit-v-t world? and what did they lose? So this was a study where we partnered with colleges around the country.

And what we wanted to find out is, is AI changing something about the way students write when they apply to college? So we look specifically at the essays that students write, these personal statement essays that maybe everybody remembers writing when you applied to college. And the question is not just, is AI making the essays better or is AI making the essays worse? But how are the essays changing?

And I think that's an important theme to keep in mind because AI is never all good or all bad. And I think, you know, if you demonize it or heroize it, you're probably too far to one side or the other. But being aware of exactly how it's changing our thinking, that is important to us. And so with, you know, a lot of time and relationship building, we were able to connect with admissions offices at these different universities. And they were willing to share with us on an anonymized base. a bunch of essays. So we ended up with actually more than a million essays, and we're able to do this study in about 370,000 of those essays. And the question we asked was based on some other work that we had done with AI in writing, and we had a suspicion. And that was this suspicion that

AI tends to make written work seem creative. And if you ask people to evaluate work that was written with AI help or written by AI, they'll say, yeah, that's really creative. But there's another thing that's happening at the same time. All that work that's judged as creative, it turns out it's quite similar. So there's something that's happening that's generally referred to as AI homogenization or AI sameness or hive mind sometimes gets thrown around. And so this was a real puzzle to us, how can you have work that's evaluated as creative, even more creative than what humans write in many cases? But at the same time, is getting more similar. Isn't creativity about distinctness? Isn't creativity about originality? And so we had a suspicion, and the suspicion is that

what AI is doing is tricking us using a cue that has worked forever and ever in human writing, which is specifically that if words are more varied, ideas tend to be more original. That's been true as long as humans have been writing, and that's a pretty well-understood connection. But we had a suspicion that AI might be breaking that link between words and ideas, that it might have the capability to generate this flood of varied words

without actually doing anything different at the conceptual level. And that is because that comes back to this question of process, right? Because when a human brain looks for words, it's looking for words using the same networks, the same semantic links that it uses to search for ideas. But AI is doing it differently. AI is searching for the next word token. it's using a stochastic sampling across a learned probability distribution,

which is fundamentally different from how brains work. And because of that difference in process, AI can divorce what's happening at the surface or word level from what's happening at the idea level much more readily than human brains can. So that was our big suspicion. So your theory going into this paper is that artificial, intelligence will use an expansive, you call it variegated vocabulary, a kind of sophisticated vocabulary to disguise the fact that across essays is actually quite homogenous. It's writing a lot of the same tropes. It's expressing a lot of the same ideas. And so even if at the word level

or even the sentence level, the writing feels varied and somewhat alive, at the higher level of ideas, there's not a lot of difference between Johnny's essay and California and Samantha's essay and New York and Jose's essay in Texas. That's your theory going into this. What in fact did you find? That's what we found. And to be fair, we had some pilot data that pointed us in that direction. So it was a guess, but it was an informed guess. But what we suspected, this break between words and ideas, that's what we found and we found it at a massive scale. And we found it everywhere we looked. So we found it separately at each of the universities that we were partnering with. We saw the same thing that after GPT arrives, late 2022 versus before, there's this very puzzling

and very consistent phenomena where the words get much more diverse and both the ideas at the sentence level and the ideas even at the whole essay level, as you mentioned, get less diverse, more similar to each other at the same time. And we did a couple more things just to make sure that it was actually GPT, or actually AI, that was responsible for this. So one of the things we did was to invite students who had written their essays before AI to come back and use AI to revise or rewrite their essays. And we saw the same pattern emerge even more strongly. And actually in those essays where we're able to directly compare human writing to AI writing, that relationship between words and ideas, which has always been there in human history, started to reverse. So the most

diverse words expressed the least diverse ideas, the least original ideas. That's so interesting. Is a fair analogy that it's a little bit like artificial intelligence can write songs that make use of every instrument, because AI can synthesize the sound of dozens of instruments in a way that no one outside of Navy Prince can actually learn to become facile with dozens of different instruments. But all of the songs are like the same chord structure.

It's all like C, G, A minor, G. It's all the same basic chord structure. So at sort of the individual like timber level, it's showing off this extraordinary, expansive, vocabulary, sound, but thematically, at the level of ideas, it's like every student essay is like playing the same song over and over and over again. I think that's great.

And here's the thing. It works, right? So I don't know enough about music to know if that would actually trick a music expert, but I do know that we had 22 creativity experts look at the essays and rate the essays creativity and the cacophony of instruments or whatever, you know, that big mixture of, in this case, semantic elements, these words, worked on, worked like scary well, tricking its creativity experts into thinking that the essays were actually more original.

All right, Adam. So, you know, students are obviously using AI more. They're using Chat, and Claude to write their halogen missions essays. These essays are creating a product that is creative enough that creativity experts can't even tell. They think it's better than maybe the typical student essay before the introduction of Chachapit. Why is this a problem? This is a problem because if what we want out of creativity is a variety of ideas just so that we can come up with solutions to problems

that haven't been solved in the old ways, if we want something genuinely new, then we need newness that's not just at the word level. We need newness that's genuinely at the ideal level. And we need to be so aware of it, because if it can trick creativity experts, what that tells us, and we also used AI to measure creativity,

and it tricked AI, AI tricked itself. Right? So what we have to know now is that if we're measuring the effect, of AI on creativity, we have to measure it in new ways. The old ways are exactly what's making, what's allowing this trick to work. We need to actually look at the differences between ideas, and we need to do that in really careful ways.

It's interesting because I think there's actually several values that an AI-written college admissions essay infringes on. one is a kind of writerly value. I think it's offensive to people who care about writing to think that artificial intelligence is doing more of our writing. It's taking it out of the domain of a human expressing to a human, a human thought,

and it's rather a human prompting an AI to express a synthetic thought to another human and passing it off as a human work. I think it offends people on that writerly level. A second value that I think, is infringed on, has to do with the nature of thinking. I've written a bit on how starting to use some of these AI tools has made me realize the degree to which writing is an act of thinking. And when I allow Claude or Chatbt to take the findings of several papers and write a paragraph

based on them, that paragraph might be a decent synthesis of several pieces of research, but it's absolutely not my paragraph. And when I look at it, or God forbid, you know, put it in one of my essays, which I do not do, I recognize upon seeing it that it's not, these aren't my thoughts, right? And so I have deprived myself of the active thinking and therefore somewhat divorced the active thinking, the active thinking from the act of writing. Adam, you're talking about a third value, which is the idea that if we allow AI to do our thinking for us, we allow a kind of tool of convergent thinking to limit the diversity of ideas that might be,

you know, more flourishing if divergent human intelligence was sort of populating that world of ideas. But I actually want you to talk specifically to these first two virtues that I talked about, the writerly virtue and the thinking virtue. In what way do you think artificial intelligence might be dangerous along those vectors? I think those are probably the biggest dangers. And the reason is we've talked before about process versus. product. One of the things that's scary to me if we're not careful is that the thing that's so

healthy about creativity, and we know creativity is healthy, it's healthy for aging well, it's healthy for dealing with emotional trauma, it's healthy for children developing cognitive flexibility. What makes it healthy is not the product. It doesn't matter if the finger painting is exhibited in the Met. It matters that it was yours, that you went through the process, you learned how to do it. It was an expression of something that was actually internal and meaningful and that you feel ownership of it. So Derek, when you talk about looking at the paragraph and it doesn't feel like yours, that's taking something from you. That's taking something that's actually very healthy away from you that's always been part of the creative process.

So if we decide that the way we're going to co-create with AI is in this kind of cognitive surrender way where we let it do the thinking, we give over the process in exchange for the product. I'm very worried about what that could mean for health and overall emotional well-being, and that has economic consequences, that has consequences for education, that has really far-reaching consequences. I really like this distinction that I haven't quite formulated for myself before our conversation right now, which is that the danger of AI in the arts is twofold. There's a danger at the level of the product because artificial intelligence

is a convergent technology and if we rely on it too much we're going to get more self-sameness in the world of art. We're going to get less diversity of ideas. But it might be even more of a problem not on the product side but on

the process side. Because even if one essay is because even if we rely on, rely on artificial intelligence to write one essay, and it's a little bit better than the essay we would have otherwise written. Over time, relying on artificial intelligence will allow our own capacities to think divergently and creatively to atrophy. And if we're walking around with atrophied minds, depending on some synthetic AI to do all of our thinking for us, well,

we've now deprived ourselves of the ability to think in the long run, and God only knows how dangerous that could be. Is it similarly dangerous? across populations, like, did you find in your research that certain types of students were affected more by these drawbacks of artificial intelligence that you're articulating? We did. So when you think about what drives homogenization, what drives the sameness in AI, one of the things that's really driving it is what goes into the data sets that these models are trained on, right?

And the content that populates those data sets tends to be pretty culturally normative. It tends to sound a lot like guys who look like me and you. And so if you, because we tend to write a lot of what gets published, right? And so that's then what's fed into the models. And there are a few other things that drive the homogenization as well, but that's part of it. So if you are somebody who doesn't look like me and you and doesn't sound like me and you, your kind of thought and your kind of expression is going to be less represented in the training data.

And that's going to mean that when these models aim for the next best word, which is what they're always trying to predict. They're going to predict that based on what you or I might write, not on what somebody whose experience is very different from ours might write. We've circled this a little bit, but, I want to jump right in and ask why AI writes like that, so to speak. And I think most people, listening or watching, will know exactly what I'm talking about. Like, AI uses M-Dashes at a level that is truly extraordinary.

There's a ton of parallelism that is, it's not X, it's Y, those type of sentences that kind of break an idea down into two. There's a lot of other sort of stylistic tells, you could say, that artificial intelligence. has. As someone who's really studied this up close, do you have a simple way of explaining why large language models like Chat Chabit and Claude have such a recognizable style of writing? One of the answers to the question is exactly as pessimistic as you might guess. the companies that are building these systems want to generate a product that is going to provide what the highest paying customers and most and the largest number of customers are going to find acceptable.

So taking a risk on something that's going to generate maybe content that isn't going to fit with, you know, the mass, with what the great mass of people want or what works in most corporate contexts is not something that they're motivated to do. So part of it is at the level of motivation. Who are, you know, who are their customers and what do they think those customers want? But part of the answer is really at the level of how these systems function, right? So these systems are functioning as probabilistic guessers, right? And the probabilistic, guess is, again, based in part or largely on what goes into their training data in the first place. But it's also based on the reinforcement that these systems get from the people

who develop them and the people who use them. So whatever it is that the great majority of people find palatable, find not worrying, not scary, not too different, find works within the context of their job, that's what's going to be favored. And so both of those factors point toward this kind of writing that seems to work well enough, even though nobody loves it. No one loves it, but no one hates it. And that's what's key, I think. I'm not trying to defend AI writing. I think it is gutless and bloodless and not particularly worth emulating. But I do think it's something you said here I think that's really important is that large language model output both has this architecture that shapes its output.

And there's a ton of post-training feedback that is looped back into the system. Reinforcement learning guided by human feedback is what it's sometimes called R-L-H-F. And if you have a lot of testers, a lot of human testers, look at a bunch of different language and say this is the kind of language that I like, this is the kind of language I don't like. It's possible that, and I'm offering an hypothesis that you can definitely shoot down, it's possible that scaled across the population, enough people will basically say, I don't mind a style that is like a street that's overlit with too many signs, right? It's like it's too clear where all the potholes are. It's too clear where everything is. It's almost too clear.

like where the crosswalks are and crosswalk coming, crosswalk coming, crosswalk coming. And in that same kind of overlit, oversigned kind of way, it seems to me the large language models are constantly overpressing their clarity, like making absolutely sure that they're offering a clear sentence to the reader that maybe if it's trying to break down a really complicated idea is quite welcome. But whenever it's like slightly more, I'm in a slightly more like artistic setting, I'm like, oh my God, this is like such, such wooden language. Is something like that, right, that the human feedback system is constantly putting its thumb on the scale of we want really, really clearly lit sentences that do not offer any possible ambiguity as to what is being said.

But then when you scale that style, you get all these m-dashes. no less parallelism and all of this like almost robotic effort to be as clear as possible. That's exactly right. And it comes from the fact that we all like to say that we that we're interested in creativity, that we like creative stuff, right? But it turns out that as a population, we don't initially. We only like it once, you know, for most of us, once we've realized other people like it or once it becomes kind of part of the zeitgeist, right? But the first time that you encounter something that's uncomfortable or unfamiliar, most people actually don't like it.

And so if you're getting this kind of feedback and you're getting it en masse, right, what it's going to favor is the overlit street. That's right. Yeah. Well, I wonder, almost arguing against the last thing that I said, what you make of stories like the one that broke two weeks ago, where the Commonwealth Short Story Prize was awarded to a short story that was almost certainly written by artificial intelligence. I mean, here we have the same way that you had creativity experts, so to speak,

grade these student essays as superior when they were inflected by artificial intelligence. Here we had, theoretically, lovers of fiction, haters of artificial intelligence, who nonetheless accidentally awarded the prestigious Commonwealth short story prize to a piece of writing that was done, it seems, now by artificial intelligence. Like, what is this telling us about either the truth of creative output or our inability sometimes to even detect the distinction between AI and human writing? Yeah, I think that's exactly where it points us. And I think what it tells us is that these tricks, these language-level tricks,

work. And that the big irony here is that if we want to be able to detect when we're being fooled, we're going to need some computational help. We actually, our human detection systems for identifying originality and language, which have always, always worked, don't work anymore. We have to recognize that. And in order to see, even to know where these pockets of homogenization live, we're going to need computational help to map that out. And that's one of the things that we're really focused on doing that we do a little bit in the research you're talking about, that we're doing in some newer research. And there's some hopeful evidence down that road. Can I talk a little bit about something there? Yeah, absolutely. Can I let me present a scenario that I think is going to be,

if not undetectable, at least very, very hard to detect. It's one thing if I just prompt ChatchBee to write a college student essay about an experience that I had at summer camp that taught me about courage and what I want from this life, right? I'm thinking of something so much generic.

That's scenario one. But scenario two is what if I say, and I'm thinking about writers who I think have quite distinct styles, write this essay in the style of someone who grew up reading Tony Morrison but realized

in midlife that they actually preferred the sentence constructions of Philip Roth. Now what I'm doing is taking this technology's ability to map meaning and language and inflecting it

with such a specific sort of recipe that I wonder how could you Adam possibly create a tool that could be precise and specific enough to measure that prompt being artificial intelligence? This is where we get to a really important point. So

it's not that the story itself isn't good. It might be outstanding. And the story that won the Commonwealth Prize is really, really good. It's that we're not talking about individual stories anymore. We're talking about what happens when you pull the stories together, right? What happens to the diversity of ideas when you're generating these stories at scale,

which is what these systems are doing? So it's very, very important to separate those two, and we're really bad at separating those two. The ideas that AI generates can be excellent. But when you put them together, it's their similarity that undermines. that undermines the possibility for generating new ideas, because it's excellent within a narrow band, right? But if you want new ideas, you have to be able to expand beyond that narrow band.

And here's another really interesting thing, I think. So we asked a question that was kind of scary to us to ask, which is, okay, so human ideas are beyond that narrow band. human ideas are more diverse than AI ideas. We're not the first lab to find that. That's pretty well established. It's not just for essays. It's not just for short stories.

It's for pretty much anything that AI and humans do when you compare them. The AI version tends to be in a more narrow band. The human variety tends to be more diverse. So the question that scared us, but we felt like we had to ask, is why is that? One reason it could be, is because AI has avoided the bad ideas that humans think of. Maybe human ideas are more diverse because humans come up with a bunch of bad ideas as well as good ideas, right?

So there's a scenario where really what we're looking at here is AI avoiding all of our stupid ideas. And that makes the band seem more narrow. That would be a plausible explanation. So we looked at a bunch of creativity tasks that humans did. We looked at our essays. We looked at several other written forms of creativity. And we asked this question, are the human ideas that are outside the AI homogenization,

what we call outside the bots, BOTS, outside the bots, are those ideas just bad ideas? And so, you know, with a little bit of a drum roll and a lot of trepidation on our part when we finally, you know, we're ready to look at the data, the answer was pretty hopeful. So actually, it turned out the quality of the writing, the quality of the creative ideas that were being generated by humans outside of the AI homogenization zone were quite, what was quite good. And in fact, when we looked at our essay data, what we saw is that those students that were thinking outside the bots were going on to achieve higher GPAs in college and actually scoring higher on tests. So it wasn't just the dummies, right? And it wasn't just the bad ideas.

There's actual value in human thinking outside the bots. And so this is, again, this is hopeful, but it's not hopeful if we give over our process to AI because then it actually flips from hopeful to scary because all of these really good ideas that are outside the bots will cease to be. If I if if AI is doing the thinking for us, all those good ideas that could lead to new innovations that could solve problems that we haven't been able to solve yet, right? Those will be the ones that disappear. It doesn't mean, and I think this comes back to something else we talked about, those ideas aren't necessarily better than AI, but they're good and AI doesn't think of them. So we need that variety.

in order for creativity to thrive. I want to ask a question, I think might either upset you or maybe rankly just a bit. You're studying what you call the ontology of creativity. And you acknowledged, I think, in my first question, that I'm not exactly sure yet what creativity really is. We're still trying to understand what creativity means, both at the output level and at the process level.

Do you think it's possible that we can learn a little bit about what creativity really is by studying large language models and artificial intelligence. Because sometimes, sometimes, despite the fact that I agree with most of what you've said here, I think, you know, AI combines this ability to remember or access a huge amount of information and synthesize novel products based on some combination or recombination of those memorized or stored ingredients. And that's just not so different than writing, then making music, then coming up with a new idea,

then coming up with a new invention. right? The process of recombination that is so necessary for innovation in human creativity is also present in the working of large language models. Is it possible that we can learn something about what creativity truly is by studying what large language models truly are? We can learn a lot about creativity, both in these systems and in ourselves, by studying how it's happening in AI, right? And so I think something you said there's really interesting, because you're absolutely right. What we do when we create is we take what we know, and in many cases, we're recombining, we're looking for ways to generate something novel

out of those recombinations, right? It is correct that AI is doing the same thing at a very high level, but where the rubber meets the road, it's doing it in a very different way. So you could say, sure, when I mix a drink, I'm mixing things, and when I invite friends from college to hang out with friends from high school, I'm mixing things. Those are two very different things. You could call them both mixing and at a high level, you're right. But the details matter a lot. And so when we come back to how brains work, when brains do that sort of thinking, that sort of combination of different meanings. And when we think about how AI is doing it by, you know, this sort of sampling over this distribution, as opposed to searching a semantic association network, right? That combination

means very different things at the detailed level. And so your point is, I think, absolutely right. There is real creativity happening in these AI systems. It's a new kind of creativity. and it forces us to look at what we mean by our own kind of creativity. So both understanding how that works and understanding how it's importantly similar in some ways and different in many ways from our creativity, I think informs our understanding on both sides. And the finding of words coming apart from ideas is a classic example, I think, because now we're going to have to really confront what do we want from our own creativity? Do we want it to sound good or do we want a diversity of ideas?

I think we want both. I want you to go one level deeper here. I would love you to explain one way as a neuroscientist. We believe that the human brain process of creativity is meaningfully distinct from the way that large language models produce that which even some creativity experts deem creative. What is one really clear distinction essentially between the way brains work and LLM's work? Brains work by putting meanings next to each other based on our experiences with the things that we interact within the world. So if you interacted with a baseball bat and that had something to do with mowing the lawn for you, right? Then those two things are a

associated in your semantic network. And we travel through the space of ideas by hopping along those associative links. We do that when we look for words and when we generate ideas. AI is doing something different. AI is putting meanings next to each other largely based on a sort of averaged semantic association across a population. So that's one thing that takes away the distinctness. But it's also searching for words much more independently from how it searches for meaning.

So it's searching for words as the next token that it's going to predict for the best sentence or the best story. And it's doing that, again, in a probabilistic way, across a learned distribution, which is very different from sort of wandering through a semantic associative network. the way that we do, the way that brains are built. I think I know what you're saying, but let me just push one more time. The same way that I remember mowing my backyard and there being a baseball bat near the yard,

thus creating a short distance for me between the concepts of mowing a yard and baseball bat, How is that idea really distinct from the fact that large language models map short versus long distances between concepts that appear more frequently and less frequently in the memory, so to speak, of all collected digital human writing? In a way, aren't we both human and large language model mapping distances between concepts? It's just that I am mapping distances between concepts in the corpus of my own individual memory, whereas large language models are mapping the distances between concepts in the collective memory, so to speak, of the entire pre-training corpus. Why is that really so distinct

as opposed to just being, I think for myself and large language models think on behalf of the collective works of all mankind? Well, let's say you're right. And if you're right, and if that's the only difference,

it's still really important when it comes to creativity because what's going to allow me to generate something different from somebody else is that unique experience, that unique associative journey, So when we look at what's going to generate that diversity of ideas, it's going to come from

those quirky, idiosyncratic ways that we connect meanings in our own lives based on our own experiences. So I think your hypothesis is going to turn out to be right at least to a certain degree, that we know because we built them, that these spaces of meanings that are sort of the underlying basis for how these systems work are based on mappings in. in terms of proximity or use in combination of these different words or these different meanings, right? But exactly how they end up being used

by these AI systems and whether that is really similar to how we search our own distinctive memories, or even how we structure our own distinctive memories. The jury is still very much out on that. And the way that the actual words versus ideas are generated, we know is quite different. I want to finish by talking about the role of AI in school. I mean, you research at Georgetown, you teach at Georgetown.

Surely you've done a lot of thinking about the fact that artificial intelligence makes cheating much easier, especially on take-home tests. But you don't seem like someone who wants to throw the baby out with the bathwater and just ban the presence of artificial intelligence entirely from education. What is your recommendation to, let's say, colleges trying to find some way to encourage, students to think for themselves, maybe use your artificial intelligence when it can be helpful, but also demonstrate their mastery and their learning rather than turn in assignments that are essentially outsourced to this averaged out Hivemind that you've been describing. How should schools think about AI? I really like the question. We're working with schools on this question, including some

admissions offices. And the answer to me is that it can't be about detecting cheating, can't be about calling AI somehow a form of cheating that we just need to try to eradicate, that whatever you think about that, that ship has sailed. So the question now is exactly the one you're asking. How do we evaluate what students are producing when we know that in all likelihood, many, if not most of them, are going to be producing it in collaboration with AI? And the answer comes back to this idea of looking at, at the distinctness of human thinking from AI homogenization.

So if you're adding what we call idea value to your interaction with AI, that will be reflected by your idea landing outside of that homogenized space, even when you interact with AI. And so if you're thinking for yourself, if you're generating something that AI alone couldn't have given me, then I should be able to see that by identifying

by mapping where your ideas land relative to the homogenized spaces that you get from AI. And the thing is, you can't see those spaces. You can't see those homogenized zones. We need some computational help to measure that. And so to me, that's really the next frontier, and that's what we've been working on. Can you leave listeners with, like, a practical rule here? Like, let's say, you know, maybe one lesson each for a student and a professor. here you have this technology, that because it is this machine of averageness in many cases,

it's going to make a lot of people's writing better. There's a lot of below average writers, like almost by definition, that's roughly half of us. It's going to make some people's writing better, but it will also, by outsourcing the thinking process, deprive you of the practice of thinking in the first place. And that's certainly bad in the long run. So I can imagine the urgency to come up. with some practical advice for students. But then also, I just think, you know, just knowing a few professors as I do, that some of them just like don't know what to do about it. They're like, they want to

ban it. They want to do the easy thing. They want to say, like, look, it's a cheating machine. We should ban it. What's one practical advice that you have for a student, practical advice you have for a professor? I'll start with the professor because I know that better. Scare him. So this is advice from Bull Durham as well. But if you, yeah, it's a great scene in that, movie. But on the first day of class, the last two years, what I've said to the students, because I still assign written papers or written grant proposals, written essays, that if what you're generating is what I could get from AI, then you have no value in the new economy, right? Because AI can do it for me much faster and cheaper than you. So why would I hire you if I could

get the same content from AI? I do think that that lands in my experience. I see that register on some faces. But I also try to give them a real genuine pitch of process matters for health. Process matters for well-being. Process matters for how you feel about your own work that you did, about your flourishing, about your mental life. This is the value of creative processes,

what's always made creativity healthy. For students, what I would say is that if what you're feeling is that other people are benefiting from using AI, other students are getting ahead of you by using AI, remember that if you're developing the capability to work with AI and still add value, then you're going to have an advantage over those people in the long run. and also that that takes practice, that there's now a new thing to learn. It's largely overlapping with what's always been important, which is to think for yourself and to learn how to write well. But learning how to do that with AI as your partner is something you should actually embrace,

something you should lean into, and something that you should practice. But you can't get sucked in to letting AI do the thinking for you. You have to hold on to that process. The three words I want to pull out from that last answer are the long run. And I think this is something that applies to both the process side and the product side. On the product side, in the long run, if we lean too heavily in artificial intelligence, which is this world homogenizing tool, we're going to generate fewer and less diverse ideas. and that's going to be bad for any domain of art

or maybe even domains of science. I've seen research suggesting that in science, because it's so easy to reference the same papers over and over and over again, you have a lot more AI written papers that are more duplicative rather than truly innovative. But then also on the process side, in the long run, as we've said, if you rely on AI to write one essay for you,

okay, maybe that doesn't have any effect on your long-term capacity for creativity. But if you develop a habit of letting AI do all the writing for you and do all the reading for you and do all the synthesizing of reading into writing, well, now you've outsourced the entire process of thought. The entire process of creativity has now been outsized to a machine such that it's a little bit like if you've gone to the gym for six months

and you've relied in a robot to lift all the weights for you, Like, your body is entirely falling apart. You haven't lifted, you know, a weight in 180 days. And I think a lot of people who are over-leaning artificial intelligence are going to feel that atrophy and process in the long run. So, I know, that's the takeaway that I really remember here, which is that maybe in the short term you don't see the effects of AI on creativity. But this is something that I think could lead to long-term atrophy, both on idea diversity

and our own capacity to be creative. So Adam Green, thank you so much. I really learned a lot from this, and I appreciate it. Thanks, Derek. It was a lot of fun. I appreciate it.