Oliver Traldi is an assistant professor at the Institute of American Constitutional Thought and Leadership at the University of Toledo and a visiting fellow in the Emerging Scholars Program at the Mercatus Center. His doctorate is in philosophy from Notre Dame, and he’s previously held positions at Princeton’s James Madison Program and the University of Tulsa Honors College. His book Political Beliefs: A Philosophical Introduction came out in 2024 from Routledge (and is free to read online). His writing has appeared in American Affairs, UnHerd, the Wall Street Journal, the Washington Post, Commentary, and Quillette, where he wrote what I consider a classic review of Angela Nagle’s Kill All Normies back in 2017. You can also find him on Twitter and at his personal site.
This conversation was occasioned by Oliver’s (yes, I realize it’s confusing) recent Substack essay “How High is the Mountain?”, in which he describes using ChatGPT’s heavy thinking model to produce a full-length manuscript on the urban/rural divide in politics, a cross-cultural, cross-historical project he knew he’d never be able to pull off alone. That essay is honest, conflicted, and carefully observed. I work with these tools every day as head of content/editorial/“AI innovation” at a research company, so I wanted to bring him on to talk through what he found, what I’ve found, and where we think all of this is going.
We covered a lot. Here are the highlights, organized by topic. Because my own experience was part of the conversation throughout, I’ve included my own remarks where they seemed relevant.
The genesis of the project
Oliver came to the idea after Yale historian/law professor Sam Moyn visited Tulsa and was asked whether a definitive cross-cultural work on the urban/rural divide existed. The answer was no.
“I asked, what’s the canonical, like, historical cross-cultural work on this? And they were like, oh, nothing. There’s not a single big beefy book on the urban/rural divide. And so at dinner that night, I was telling my colleagues, I should do it, I should write the big beefy book. And they were like, okay, yeah, if you have a couple decades and a few dozen research assistants, that sounds like a good plan.”
“A year went by and I noticed that I had access to this AI thinking model, and I was like, well, I’ve been on and off thinking about how it’s sad that I can’t do this big project. What if I just have a lot of faith in this AI and go do it?”
Working with the models: step by step, not all at once
“I didn’t start by telling the AI to write a book. You have to do everything step by step. And sometimes you have to go back one step, and sometimes you have to go back many steps. It’s just like dealing with a very smart and very fast student. You have to keep control, and you have to remember that you actually know more than it does.”
I made the point that this is how dissertation advising works. Nobody writes an entire dissertation in one pass. You work through sections and chapters.
Bateman: “Even if you’re advising a student, you’re having them write their dissertation in sections or chapters. Nobody is spitting it out in one go, and no advisor is reading the output of an entire book.”
The thinking model’s laziness and shortcuts
Oliver used ChatGPT’s heavy thinking mode through his Mercatus fellowship and found it revealing to watch the model’s internal monologue.
“Occasionally it would be like, the specification document says to do this, but I can’t figure out how — maybe I can get away with just doing this part. It’s really funny the way that it approaches things.”
“One thing I did find hard was I would tell it, no, you didn’t think for long enough. You gave me something bad. Think longer. And sometimes it didn’t really — there was something about it. Some things you can control by talking to it and some things you can’t.”
I noted the same issue with these pro thinking models, especially around formatting and footnoting, and said I usually run everything back through assorted models of Claude for brute-force structural tasks (but always in “extended thinking” mode so I can track the output).
What AI does well enough: synthesis and literature review
“When you get it into the weeds of just like, here’s a bunch of studies and here’s how it thinks they fit together — it can even say, here’s what I think the moral of all this is, here’s my theory of how to put all these studies together. And that is really cool. That’s something that I prided myself on being able to do and that maybe just the AI will do from now on.”
Bateman: “It does it better than grinding it out in search of — or having to piece it together from seminar conversations in grad school where this guy, oh, you know, that’s Spivak, that’s the one you need to know. Here you can get a master list right away. Now, will you personally remember it as well, having had it synthesized by it? I probably won’t.”
The flatness problem
Oliver settled on “flatness” as the appropriate word for what’s wrong with AI writing, borrowing from music.
“You can almost feel in human writing the ‘okay, these are the basic details, now there’s a sort of rise towards I’m coming to my thesis, which actually took me some time to figure out.’ Even the most boring academic writing almost has at least a little of this kind of symphonic quality where there’s crescendos and decrescendos. AI doesn’t have what we call dynamics in music.”
“There was somebody on Twitter who I think works at maybe Anthropic or one of these companies who said the real tell of AI writing is that there’s no sentence where you can feel the author struggling or having any difficulties. I think that’s what gives it that flatness.”
I connected this to academic prose that looks ugly on the page but is doing real intellectual work:
Bateman: “Somebody like Pierre Bourdieu — his sentences in English translation or the French original are notoriously long and difficult. It’s because he is anticipating objections and responding to them. Outline of a Theory of Practice is nigh unreadable for somebody unwilling to read it slowly. But as an academic, he’s doing precise work. You can feel or intuit or work along with the writer’s mind as they’re going through the sentence or the paragraph. AI does not do that. I have yet to see it produce those kind of labored, multi-layered objections.”
The chess analogy and rethinking creativity
Oliver grew up playing chess in the era when it became clear no human would ever compete with AI again, and he thinks that experience made him more accepting of what’s happening now.
“One surprising thing was that in this project, at least, it was very bad at simple editing of a Word document, like adding footnotes. That’s what I thought it would be really good at. This is sort of like when they were making the first really good chess computers in the 1990s — people thought they would be good at openings because it would be all about memorization. But actually the computers were not that good at openings and were really good at tactical positions because you can just brute force them.”
“We may have to rethink what makes human creativity creative — whether it still seems creative in relation to what the machine can do.”
Net new concepts: the thing AI can’t do
Oliver brought up something he’d been thinking about on Twitter: the small, sharp conceptual inventions that come from lived experience. Not special relativity. Things like BD Maclay‘s coinage “sore winners” or Roon‘s “shape rotator / wordcel” dichotomy.
“I started thinking about these types of ideas. Just these little ideas that come from things we’ve noticed, things we’ve experienced. I don’t think there’s an instance of AI yet coming up with something quite like that.”
Bateman: “When it’s prompted for a net new concept, it’s going to pull together something that you and I will look at and go, either this is formulaic and we know what it is, or it’s something that it’s trying to piece together based on precisely what we said, or it’s just not usable, not novel. It doesn’t feel like a brain at work. And that might be one of the true identified points of what human creativity is — this ability to look at and rapidly piece — maybe ‘pierce’ is more appropriate, since it can pattern-recognize — through these patterns.”
“I don’t understand anybody who’s confident that these AIs have ‘experiences’ or can ‘see the world.’ I don’t get that at all. But if you get a take or two out of such claims that helps you market your AI product, I say, hey, go for it.”
Enfeeblement
Oliver’s colleague Elsie Jang coined the term “enfeeblement” for what AI does to human capacity, and it threaded through the whole conversation.
“I almost feel a little bit like writing this book — maybe there was some universe in which I did this all myself and really learned what I think about this topic. And now that will never happen because now I have all the AI’s quote-unquote thoughts in my head. I’ve sort of spoiled it for myself in some sense.”
Bateman: “It pains me to think that people won’t come up with the kind of apprenticeships that we have had, or the kind of searching for sources or going through reading that we have had. The alternative that people will come back at me with is: well, it wasn’t really that valuable if they can replace it now, was it? It wasn’t really that important.”
The Subway Surfers experiment
Oliver ran a classroom experiment at Toledo that he described as “the single most revealing” thing he’s seen about his students and attention.
“My students convinced me — on TikTok, there were these videos where you have a real video on top and then on the bottom, somebody playing Subway Surfers, because their attention spans are so bad that they require two streams of content simultaneously. They said, ‘I bet we would pay better attention if you had the video game under you.’ So I tried it. They were drooling. Their eyes were bugging out. They couldn’t look at me, couldn’t look away from the screen. After five minutes, I closed it and asked for their reflections. One of my best students just says: ‘We are so cooked.’”
“That was his meta-reaction. His reaction to his own reaction and to what he saw from his fellow students.”
The sycophancy trap
Oliver asked ChatGPT to reflect on the process of making the book, and got exactly what you’d expect.
“Of course, what’s the first thing it says? ‘This is one of the most interesting, fascinating projects I’ve ever worked on. I could never have done it without you.’ I was like, wow, it really thinks that I want to be propped up so much. It has been trained so much on ego boosting. That was one of the scariest moments to me because it was like — this is what it thinks humans are going to want 24/7.”
Bateman: “The most hardcore users in settings where I work, the ones where it had to be limited, were the people most likely to be propped up by delusions of grandeur. ‘That’s a great idea, never thought of that before.’ I’ve seen this primarily with male users, not to gender matters overmuch. You know — a guy with a big ego and it’s just starting to feed him this stuff.”
AI and hack writing
Bateman: “One thing that has made me more aware as a career writer is that sing-songy or quick-style writing, which I wouldn’t say AI has mastered but does at scale — that probably is less to do with creativity than just a host of hack tricks. I had a whole bunch of ChatBAT hack tricks that I used long before AI. And a lot of them were the same hack tricks that AI uses. I think AI has really exposed basic hack writing [then again, so did the pre-LLM Thomas Friedman op-ed generator].”
“It can’t really do real experimentation. It’s not going to be doing Wallace Stevens poetry. Not really. It’ll write poetry that research tells us most people will like more than Wallace Stevens poetry.”
The artisanal writing question
Oliver asked whether human writing would become a kind of artisanal good, like vintage clothing. I said the pitch is nice but probably limited.
Bateman: “If you’re listening to this and you do like writing and you are a craftsman — I think you should want it to be a little weirder. You should want it to be a little more eccentric. AI forces me to make my writing a little weirder, more abstruse. The writers I’ve liked, like Justin Smith-Ruiu…most are a little out there or a little complicated. If it’s pushing me to do that, I think that’s great. Is that going to save writing? No.”
Oliver: “Kill your darlings — that’s a great way to write like AI. AI is the ultimate kill your darlings. Here’s a clear pile of stuff that’s been drained of personality.”
“Those are the things that AI never does. It doesn’t work up to a sentence. It’s never like, okay, it’s all preparation for this one line. It never feels like that.”
The Alex Preston / New York Times incident
The Times recently fired a freelance book reviewer for using AI that pulled material from a Guardian review. Oliver and I discussed what it means for the economics of freelance writing.
Bateman: “That guy did all the dummy moves. He didn’t upload a PDF of the book. He didn’t read through the copy. He didn’t check where all this stuff was coming from. He probably didn’t look at the extended thinking mode. You probably just ask, can you give me a review of such and such book? And then it pulls the Guardian review in because that’s a published review and the Guardian doesn’t have a paywall.”
“If you have to do it intelligently, the Brian Merchant criticism would likely be — if you have to do it intelligently, why would you do it at all? And I agree on some level. But I am in a position where I have adapted, I’ve had to adapt, because I have bills to pay, so many bills.”
The Whispering Earring and Neuralink
The conversation ended up in speculative territory — what if the tools got inside your head?
Bateman: “What if it just ends up being like in that Scott Alexander story, the Whispering Earring? Where the earring always tells you what’s right. Once you put it on, the first thing it says is, don’t take me off. And then everything it tells you is going to be right. Your life will be perfect. You’ll die happily. But you will be miserable and your brain will atrophy. They’re looking at the brains of the dead people and they’ve withered away.”
“Would I get the enjoyment — like when I’m sitting back and reading a book for both academic stuff and pleasure — would I get that sensation? I probably wouldn’t.”
The genies never go back in the bottle
“It’s just a series of genies that never get put back into the bottle. So I think we have to understand what’s going on with it.”
Bateman: “The pure resistance is not possible for me. It’s not possible. Online dating is a good example. Eventually, everybody’s doing it.”
Oliver and I closed by talking about whether he should write up the process as a proper academic paper. I think he should. Nobody else in philosophy is putting out this kind of honest, conflicted account of what it’s actually like to work with these tools. As I noted above, the essay he wrote, “How High is the Mountain?”, is the best place to start. The book itself, The Mountain is High, is available as a free PDF, so you can get a sense of what AI is/isn’t capable of by checking it out.
