Maybe it’s good, maybe it’s bad, who knows
I have a strange attitude towards AI. Strange to myself, I mean. But I suspect it’s actually how I am in general, and maybe you too, and not just about AI but about everything. My attitude is something like this:
The way I use AI is correct
The way most others use AI is wrong
I am the only person who is avoiding the trappings created by using AI for research
I notice this because I notice on the one hand optimism towards AI when I think about how it affects me, and I notice pessimism towards AI as I start working through a series of deductions about AI and its causal effect on one’s allocation of time in doing research. And I constantly end up a bit muddled. On the one hand, I am optimistic but that optimism is almost always about myself. And on the other hand, I am pessimistic and that pessimism is more general, towards some huge looming equilibrium.
I’ll share an anecdote. I have with four sets of coauthors each their own four very large projects that we just keep seeming to be stuck on. We aren’t stuck on them because for some technical reason. We are stuck on them primarily because we do not simply devote the time collectively to finish them. That is because what it will take to finish them is 50+ hour work weeks for about two to three weeks without interruption, and if we each spent 50+ hours doing nothing but this project, then we’d be done.
But for at least three of those four projects, that would literally be “writing”. Because the work is done, or we think it’s done. But see, that’s just it. When I look back over my entire 20 year career as a PhD economist, my work has never been neatly separated where at the end I am writing and only writing. Almost always my work has been that during the writing, some new problem is found, we work on that problem, some new stuff is discovered, the project changes — sometimes a little, sometimes a lot — and then we keep writing, and then through this constant spiral that is getting tighter and tighter, we eventually finish.
So in a sense, yes — yes, in a sense the problem is that I need time. We need to all be in the same physical location, probably, for 2 to 3 weeks. Not on Zoom where we finally get online for 1 hour, we inevitably talk for a second, we are dealing with kid stuff. No, we just need to be in person, together, for 2 to 3 weeks and if we can be together for 2 to 3 weeks, not only will that entire last part of the marathon be completed, but that the work will be improved. It won’t just be written — it will be substantially improved and written.
Well, what I tell people all the time is that the proper framework for thinking about AI’s effect on one’s productivity is not to just assert it’s some crank that raises your marginal product. Maybe all technology is what I’m about to say, and AI just makes it more clear that it is this because it’s the last technology that has had this effect on cognitive output. But the proper framework for thinking about AI’s effect on one’s productivity is time use. It's think it’s specifically some coherent combination of two distinct strands in Gary Becker’s work. His work on the allocation of time use and his work on human capital.
Well, if AI allows you to complete cognitive tasks without using time — and while people often object to my illustration of this, the fact is one can zero shot “submission ready manuscripts” in less than an hour, hands completely off the steering wheel. And any time someone says that it is not possible, I show them it is possible and they are stunned and almost always it means that the reason they thought it could not was because of user error. Or because they have some threshold for reservation quality that they have in mind to which I always ask — where exactly is this particular zero shot paper in the distribution of all empirical papers written in 2024 globally? Are people really saying that the fully automated AI empirical research papers, perhaps steered by someone who really does know how to steer it, saying that they are the absolute min in that distribution? That is an empirical question.
But sorry I’m off track. Let me get back to what I was saying. If AI allows you to complete cognitive tasks without using time, then what will I be doing in those 2-3 weeks of face time together? Will I even need it? I mean, AI can write the paper for me, can’t it? Yes, but that wasn’t exactly the reason we needed 2-3 weeks together in the first place. I mean yes we needed to finish the project, and we need 2-3 weeks together to finish the project, and we think we are at the end, we think we need to write. But the thing is, we need to write and we need to talk about the writing and we need to see things at the same time and we need to talk as we see them at the same time, and we have to tolerate each other’s efforts to understand it at the same time, and we need to listen to one another and understand one another because if we don’t listen and we don’t understand we won’t be able to finish and something is blocking that process, and that blocking can only be overcome through a particular kind of time use deposited together and being physically together seems to make it happen, but not always, and so forth.
I mean am I the only one who is like this? Is everyone else just so efficient with their time that they don’t need to somehow simply be together, in closed quarters, doors closed, door dash delivered, white boards and blue markers and feet on the table and going home to wake up to texts from your coauthor that goes something like “I found more problems.” And through some miracle of a weekend, problems unearth solutions that radically overturn the previous work and lead to discoveries that could only have been discovered by confronting problems and driving through them not around them?
So then back to AI. AI allows you to complete cognitive tasks without the use of human time. But what are these tasks exactly? Are they “writing”? What is writing in the hands of human researchers? Is writing in the hands of human researchers the same thing as writing in the hands of AI agents? I mean, that’s the puzzle — it is and it isn’t. It is in the sense that AI agents and human researchers could frankly end up in the limit writing the identical manuscripts and yet because the process that the human must undergo in order to write that manuscript required a process involving one’s time inputs that it is actually impossible to write that manuscript without the use of time.
What is this time use exactly? Is it measured in minutes? Yes it is measured in minutes. So that can’t be it. I think the time use is something to do with the object of time itself. I say all this because I think I notice that if I think of “writing” as simply this task of tapping on a keyboard, then AI is a perfect substitute for it and all complaints about AI written manuscripts comes down to judging the quality. Who’s writing is better?
But the strange thing for me is that “writing” actually seems to be about that physical togetherness, ordering pizza, one coauthor going to bed at 10pm, the other coauthor staying up until 2am, texting the first coauthor they found a problem at 2:30am, the first coauthor waking up at 7am, getting the text, looking at the new results, having their coffee, running on the treadmill, and then coming in together to talk about it. I mean am I really the only one for whom “let’s finish this paper” means “let’s write this paper” means that exact process which seems to have nothing to do with actual writing and yet is what seems to be happening in the “writing stage”?
So back to my optimism/pessimism. I went to a university on Friday last week as a kind of consultant. I was asked to come and spend half a day with a department to talk with them about AI and their program. It was amazing, and I honestly wondered afterwards “maybe I should tell people to have me come out and I could do more of this”. But you know what one person told me made the experience valuable? They told me that one thing that made it valuable was that because I came out, they all got together in the same room, and for half a day they interacted with one another around this particular topic that otherwise they never had the chance to talk about because too many other things competed on their time.
So was that time? Yes it was time. But it’s like the time is spent on a task that almost cannot even be described easily. It isn’t learning time and yet the collective time spent together creates learning.
I am optimistic about AI for empirical research. I told people at the consulting thing that I really felt like you had to try to ignore prophets telling you what the future is going to be because those prophets do not know anymore than anyone else what the future is going to be. No one knows the future. All discussions about the effect that AI is going to have on this or that are at this exact moment based on private models, rarely written down, and rarely worked through. They are based on “folklore models”. Which isn’t to say those aren’t helpful, but I think it is important to remind ourselves that all models are wrong, but only some models are useful. So we have to ask ourselves why are we telling ourselves this model? Why are we believing this model? What exactly are we using this model for? What do we need right now? What needs to be done right now? Whatever needs to be done right now is the thing that makes a model useful. Does this model we hold up as some kind of inevitable fact, as if it’s simply too obvious to even be debated, help us achieve to the best of our ability what needs to be achieved right now? If the answer is yes, then it’s useful and if the answer is no, it is not useful, and we abandon the useless ones and we keep the useful ones.
So then what exactly is it that needs to be achieved right now? Well for you reading this, and me, it’s our jobs. We have jobs to do. We have to finish these papers. We have to teach this class. We have to mentor these students. And it is so tempting to get pulled into a kind of hand wringing about the inevitability of the future that we spend less time on the things at hand. Which is mentoring, finishing and preparation.
Well, I guess I am optimistic when I am thinking about the things at hand. I have to finish this paper, I have to fix this class, I have to help someone. And AI helps me do that, and so I use it, and I don’t when it doesn’t, and I just try my best to not start connecting the dots. I try not to connect the dots. I try rather to more or less flow through the dots, live in the dots now, and not spend any time drawing out all these necessary deductions.
So, am I optimistic? Yes. I am optimistic because I have tasks at hand, and I have to get them done. I still have to get physically together with people, and I’m realizing that more and more. And so I have to make that happen. And AI isn’t going to be able to make that happen. I have to stop what I’m doing, turn things off, and just be with my coauthors, my students, my colleagues, what have you, and simply be there doing it.
But maybe when we are together, then I just let myself do whatever I want to do to make what needs to be done get done. Which for me now always means using agents. And maybe if I do that, I can and will trace out a story where the long run equilibrium looks unavoidable and that’s where humans are not needed to be in the loop at all.
I just simply think that that is a story. That’s a story about the future. And just like all models are wrong but some are useful, all stories are wrong, but some are useful. And if the story ends up making me not do the things I have to do now, then it’s not useful, no matter how true it seems to be, and it’s on me to say right now “This is not a useful story, because I have work to do, and so right now I am not going to consider it because in considering it, I seem to not be doing the work that needs to be done.”
There is a story called the Farmer and his Horse that goes like this. A farmer’s horse one day ran away and the neighbors saw it and told the farmer what bad luck because the farmer needed that horse for his livelihood. The farmer said “maybe it’s good, maybe it’s bad. Who knows?”
Then the horse galloped back one day with four other wild horses and suddenly the farmer was rich because now he had five horses, not just one. To which the neighbors said “what great luck! Now you have even more horses!” To which the farmer said “maybe it’s good, maybe it’s bad. Who knows?”
But then one day his young son tried to break one of the wild stallions, was thrown off, and broke his leg. And how the farmer’s son was injured and the farmer couldn’t get the work done he needed to get done around the farm. The neighbors saw this and said, “Oh no! What bad luck!” And the farmer said, again, “Maybe it’s good, maybe it’s bad. Who knows?”
But then yet another event happened. The local warlord came to the village and conscripted all of the young men to go off and fight in a war. They all leave … except for the farmer’s son who cannot fight because his leg is broken. And the neighbors look at the farmer and grimace, “What good fortune has befell you.” To which the farmer says, “Maybe it’s good maybe it’s bad. Who knows?”
I think of this a lot ever since my former colleague, Rebecca Thornton, told it to me years ago. For someone like me, who lives in his mind, and who has a constant flow of compulsive worries and fears and anxieties, it’s a very interesting story because it challenges the very thing that I face all the time which is that my thoughts about the world are the world. And that my thoughts about the world are always somehow mixing several things at once: description about the world, description about inevitable processes about the world, predictions about the future, predictions about counterfactuals, explanations of the present in terms of the past, and normative rankings of everything. But all masquerading as something so obvious as to be ridiculous to think otherwise.
I think it’s really not clear what effect AI is going to have one you or me or society. And if you don’t want to use AI for empirical research, that’s absolutely your prerogative. And that’s it — full stop. There’s nothing left to say. If you do not want to use it, you should not use it. The right reason to do something is because you want to do it. Not because you have been conscripted to do it. You do it because you want to do it, and you don’t do it because you don’t want to do it, and you just own that. And when it seems like AI is taking you away from the thing you love, you just have to go through that too. And no one else is going to be able to do that for you (or me), and no one is going to make it easier.
I am optimistic about AI on me because I am optimistic about me and my journey. I am optimistic about my competency and my wisdom and my goodness in such a way and to such a degree that to be anything other than optimistic about my ability to navigate my historical life, and that I am perfectly capable of being the one in charge of it, would simply be an error on my part. Will my decisions be mistakes? Maybe they’ll be good, maybe they’ll be bad, who knows. But for now, I have work to do, and I need to do that work, and do it to the best of my ability, and for me, that’s going to involve continuing to try and understand.
Scott's Mixtape Substack is a reader-supported publication. To receive new posts and support my work, consider becoming a free or paid subscriber.
© 2026 scott cunningham
n/a