What Makes a Job Automatable? – with Bouke Klein Teeselink

Danny Buerkli: My guest is Bouke Klein Teeselink. Bouke is an associate professor in economics at King’s College London and chief economist at the AI Objectives Institute. Bouke, welcome.

Bouke Klein Teeselink: Thank you very much. Pleasure to be here.

Danny: Thanks for being here. Bouke, we’re recording this in July 2026. At present, what do we know about the labor market effects of AI?

Bouke: Well, that depends a little bit on who you ask — there’s clearly quite a bit of disagreement here. So I’m going to give a few statements at differing levels of certainty. Based on my own research, and on the Canaries research in the US, I’m reasonably certain that occupations and firms that were more exposed to AI have seen a decrease in hiring and a decrease in employment. That’s a statement I’m willing to make and that I’m really quite sure of.

Now, that directly raises the question of whether that is actually caused by AI, and that’s something I’m slightly less certain about. There’s quite a bit of evidence consistent with it, but there are two issues. One is that we basically had one big shock, the introduction of ChatGPT, which means anything else that happened around the same time could have caused the same thing. The second is that there are many variables that correlate with AI exposure — it could be interest rate exposure, it could be work-from-home exposure — that could then potentially explain the results.

On top of that, there’s a third explanation that is consistent with AI being a major factor, but where it isn’t automation that we’re looking at. That’s my preferred explanation at the moment. The reason I believe AI has something to do with it is the timing: most of the effects I’m seeing, and that others are seeing, quite clearly start right after the introduction of ChatGPT. That’s inconsistent with the work-from-home hypothesis. In addition, AI exposure doesn’t correlate that much with interest rate exposure, so that doesn’t seem like a very likely explanation either. And we see it in particular in certain fields where there’s a lot of public debate or public fear about AI playing a role.

So, okay: I think AI has something to do with it, or at least the data is very consistent with that hypothesis. Now, is this automation? Actually, I don’t think it is. The data I’ve been looking at, and that others have been looking at, runs until about June or September 2025. I really don’t think we were automating many jobs until that point. The technology just wasn’t there yet. Sure, people were using ChatGPT to speed up their coding or to do some of their admin work, but that’s just not enough to automate a job.

What I think is happening is that this is an uncertainty story. If you are a firm and you’re highly exposed to AI — let’s say you’re a software firm — you just don’t know what the technology can do for you a year from now. We’ve all seen these graphs of the development of capabilities over time, which is completely wild. So right now you think, okay, AI can do software tasks of sixteen hours; if we extrapolate a year into the future, it can do software tasks of three months, or a year, or whatever. Planning for that is really, really difficult. You need to figure out how many software engineers you are going to need, and if you’re hugely uncertain about that, it is actually quite rational to just defer hiring until either that uncertainty has resolved to a degree, or you really run into bottlenecks selling the product you’re producing.

And I have some evidence consistent with that. I’ve looked at AI exposure and job postings in many different countries, and broadly across the board I find this reduction in new job postings in more AI-exposed occupations, in the vast majority of countries. But the effects are bigger in countries with really strict labor laws, and that’s very consistent with this kind of uncertainty explanation. Because if it’s expensive to get rid of people and you’re unsure how many people you’re going to need, of course you’re going to defer hiring. If you’re in Germany or France and it’s going to be hugely expensive to fire a software engineer a year from now, and you don’t know whether you’ll need more or fewer, you’re just not going to hire. So that’s my broad interpretation of what we learn from these AI exposure studies. But even this is a hypothesis, one that hasn’t been tested extensively yet.

Danny: I agree. With the introduction of ChatGPT it would seem wildly implausible that you would already see employment effects caused by automation, or by sufficient augmentation that it would reduce the need for more employment. It would seem that now we may be slowly but surely entering into territory where it becomes vastly more plausible. What would you look for in the data to discern when we’re starting to tip from anticipation effects, as you’re describing, into actual displacement?

Bouke: I think we need two things. One is that exposure is not adoption, and we don’t have amazing firm-level adoption data. I’ve been working on this a little bit with Bharat Chandar from Stanford, trying to infer adoption at the firm level from job ads. Others, like Hosseini and Lichtinger, have tried to do something similar for the US. But even there, the moment a firm starts posting jobs is a proxy for adoption, and it’s not the kind of adoption we want. Ideally we’d actually know when firms start integrating this technology into their workflow, which most of the time we just don’t know. So we want this for a large panel of firms, and we want to know when.

And then on top of that, the introduction of agentic coding was a complete game changer. My job has radically changed since the introduction of Claude Code — actually much more radically than with the introduction of ChatGPT — in the sense that I don’t write any code anymore. That used to be 50% of what I do. That’s a completely radical change. And all the AI exposure measures that most people use don’t take those capabilities into account; they’re sort of fixed on 2024, on GPT-4. We need to know what the technology can do now, across all these different occupations. And then we want to see who’s actually integrating that into their systems. That would allow us to study what happens to employment in those firms. So we’re trying to get there.

Danny: But we’re still pretty far away from having a conclusive answer. You’ve done some work on measuring potential future exposure — we’ll get to that in a second. How has your own hiring behavior changed? Have you changed the number or the type of RAs that you hire?

Bouke: Yeah, this is something I’ve been struggling with a little bit. Any task that I used to use RAs for, I don’t need RAs for anymore. That’s really dumb, right? I don’t need someone to check my code. I don’t need someone to do a literature review for me, or find papers. AI is going to do that much quicker and much better. So the traditional RA tasks are over. That means two things can happen. Either I’m just not going to hire RAs anymore — which, in all honesty, is a little bit how I feel about undergrad RAs, because they’re in a sense just not ready to do more — or you give your RAs much bigger tasks.

My RAs also have access to frontier AI. I’m currently working with an RA, Jakob Schell, who also works with Windfall, and I’m giving him basically a whole paper: this is the goal of what we want to do here, you can lead it. There’s no way I would have done that before. I would have done that with a second- or third-year or final-year PhD student, not with an RA. And that’s the hopeful part here, because it’s not just that I can do my old tasks with an RA — my RA can now do much more. So if the RA is really good, which Jakob is, it’s incredible to have RAs. I can do a whole new project because of an RA, which I just don’t have time for otherwise. If the RA is not that good, though, it’s completely useless. So there is a bifurcation, where the quality of the RA matters much more, because the grunt tasks that were not that hard are over.

Danny: Whereas hiring a really good RA, or working with a really good PhD student — the returns are actually going up. There are a couple of things going on there. One is the impact of expertise, which you’ve also worked on. It depends whether we’re automating low-expertise or high-expertise tasks out of the bundle that constitutes a job. In this example, what is left are some high-expertise tasks — but, as you’ve just said, that starts to exclude certain people who, for whatever reason, cannot do those. And then for them this is a binary thing: they just fall out of this particular labor market, as it were.

Bouke: Yeah, I think that’s right, and there are two things to consider. I’ve done research on the importance of expertise. All these exposure measures, including my own, tell you how many tasks in a job can be done by AI. But as it turns out — and this is what David Autor and Thompson pointed out — it matters a lot not just how many, but which tasks. It matters whether it’s the high- or the low-expertise tasks: if the high-expertise tasks are automated, the job becomes easier, more people can do it, and you get a decrease in wages. Whereas if the low-expertise tasks are automated, the job becomes harder, fewer people can do it, and you get an increase in wages. We’re already seeing evidence of that in wage data across several economies.

What’s important, though, is that we’re simplifying a little, because what matters is not just the expertise level of the job, but also the expertise level of the people who could do the job. And both of those change. The threshold for being an RA for me has clearly gone up, because the easy tasks are basically done by AI now. Therefore I require a higher expertise level for an RA, and that also means I’m willing to pay higher wages for an RA. Fewer people can do it, so there is a decrease in supply — previously there was basically an infinite supply of RAs, and now that’s just not true, because most people can’t do it.

But at the same time there is an augmentation of the supply itself. A good RA now has much higher effective expertise than they had two years ago, let alone five years ago. So the AI changes the expertise level of the supply too, and that’s super heterogeneous. We’re at a point where it seems to be especially people who are good at something who become hugely more productive with AI, whereas people who are reasonably mediocre don’t gain all that much. So again there seems to be this bifurcation in the labor market, where you might just get a huge polarization. Because if it makes good people better at what they do, and it doesn’t do as much for people who aren’t that great — this is, again, skill-biased technological change.

Danny: And this story obviously holds true even if we fix the capabilities of AI systems at the level they’re at today. The obvious question is — and I think the research assistant example is a good one — what makes us confident that your average highly talented RA will still be able to contribute on the margin in, say, two or five years from now?

Bouke: Well, you can stretch that argument a little bit further: what makes me confident that I have something to contribute five years from now? In all honesty, I can now one-shot a paper of a decent first-year PhD student. If I give Claude Code — and I let it run on dangerously-skip-permissions — or Codex, it doesn’t quite matter which model, and I say, here’s an idea, or not even that, just: give me ten good ideas, run with the one you can do best. Even if we fully take taste out, where I do nothing except start the prompt, it does a reasonable job at this. Not an incredible job, but a reasonable job — really at a level that’s possible for an okay first-year PhD student.

If we’re there now, then given the level of advances we’re looking at, a year from now that might very well be a final-year PhD student, which means two or three years from now it would be me. And two or three years after that it would be a Nobel Prize winner. Then it becomes a question of what we still have to contribute. That’s a huge open question, and I could see it go in two directions. One where it is not that much — I’ll come back to that, because I have some ideas about what I’d be doing then. And another where our job just changes a lot.

It might be that we still need to take initiative, or that we’re still the purveyor of ideas, or that we still have to differentiate between — this AI wrote 17 million papers, which ones are good? And which ones are good is not just a quality judgment, which for all we know AI can do really well once it gets even better. There is also a value judgment there, and there’s a sense in which that value judgment isn’t fully automatable, because we have to decide what we value. AI might be able to predict that, but it’s our decision. Insofar as that’s true, maybe that’s what we’d be doing.

You can imagine a world where we have to make policy decisions based on economics research. AI has automated all the research — but that’s a lot of research, an enormous amount. We might just need interpreters. Maybe we just need people who translate what all the research AI has done into: okay, this is what it means for people, this is how we should go forward.

Danny: How is that an internally consistent story? If a is possible, why is b not also possible?

Bouke: For the value judgment, I think it’s not possible, because it is something we decide ourselves. If I have to make a choice between society A and society B, that’s not a prediction problem. In a sense that’s just my own utility function — I decide which I prefer.

Danny: But can I not turn that into a prediction problem by having an AI predict what your value judgment would be? It might be pretty good at that.

Bouke: It would be great if AI gets much better at those value judgments, because that’s a precondition for having personalized agents, or even political agents, that act on our behalf. It needs to understand our values. I don’t know if we get there, because it needs a lot of information about me — and I might change. This is a little bit like the old Hayekian problem of information. In many ways AI is incredible at aggregating information, but some of this information just doesn’t exist, because it lives in my head. My preferences live in my head. They’re revealed in my choices, so there is a sense in which you can predict them. But if my preferences change without any new choices, there’s just no way of extracting that information. And in particular, if I act less in market settings, there are also far fewer opportunities to learn from my revealed preferences. So insofar as that information doesn’t exist, it is unlearnable. There are edge cases where that value judgment isn’t automatable.

Now, we can probably get very far. If we get to the level I just described, we clearly don’t need the number of researchers we currently have. I’m sort of on board with the fact that I probably might not be doing research in five or ten years. We gradually see it now: what constitutes research is just changing a lot. At the moment we’re at a stage where coming up with new ideas is a super high-value thing to do, because the execution cost has gone to zero. There’s a complements-versus-substitutes question: the cost of execution goes to zero, therefore everything that’s a complement to execution goes up in value — which at the moment is coming up with good ideas, deciding what is a good idea, things like that.

Danny: It’s not clear to me, indeed, that AI can’t do those things. It’s not obvious that there is a fundamental barrier. People love saying that AI cannot come up with new ideas, but to the extent that innovation is combinatorial, it’s not obvious at all that that’s true.

Bouke: No, I think it’s very obvious that it can come up with new ideas. Basically all my ideas are combinatorial — maybe that’s just my own lack of creativity, and I’m fine with that. But if it can solve frontier math problems by combining different strands of mathematics that we haven’t combined yet, surely it can come up with loads and loads of new research ideas, whether that’s in the social sciences, by drawing insights from other fields, or by bringing different fields in economics together. I expect huge gains there, in ways that we don’t fully understand at the moment.

I will say that some problems aren’t combinatorial, and it’s less clear that AI — even AI a few years from now — would be very good at that. That’s a huge open question: how far it can do paradigm shifts, compared to paradigm exploitation. We don’t know. But I think there’s so much to be gained from paradigm exploitation, which includes all the combinatorial things across our current paradigms in different fields. The fraction of people who can do genuinely paradigm-shifting research is super small — a really, really tiny fraction. And maybe we just devote much more resources to it. Maybe a lot of people like me, who have a 1% chance of coming up with that kind of thing, would be working on it, because the value goes up so much: once you discover that new paradigm, you throw all the compute at it.

Danny: We just take more shots on goal.

Bouke: Exactly. Maybe we’d all be doing moonshot stuff, insofar as AI can’t do the paradigm exploration. But we don’t know. There have been so many emerging capabilities that we couldn’t predict. We couldn’t do maths three years ago; now it’s incredible at maths. I think everyone should be more blown away by what this technology can do. Just the fact that — disregarding all post-training — what we got from next-word prediction is completely remarkable. There’s absolutely no way that any reasonable person would have predicted this.

Danny: Really?

Bouke: If anyone read the transformer paper in 2017, there’s just no way that even if someone had understood, hey, we get incredible next-word prediction from this, that you get something that resembles intelligence. I would have guessed we get really good translations.

Danny: I think that was the intention, if I’m not mistaken.

Bouke: Right. So you get all these emerging capabilities that are really quite incredible, and we should be blown away by that more. And it’s clear that every time we say, oh, AI can’t do x — that seems to be our role as humans.

Danny: And a year later, the AI can do it. A year ago we all said AI can’t do taste. Yes, it can. I was going to ask about taste — you mentioned taste before. For a while this was the favorite concept to reach for when trying to describe the ineffable quality that AI can’t quite grasp, which means knowing what is worth researching, what is a good research question, for instance. But it’s not obvious to me what that is, and it seems the models have gotten massively better at whatever that taste quality looks like.

Bouke: One way to test this — I haven’t done it much — is to feed it ten research ideas that I think have a clear ranking, or even that don’t have a clear ranking but where I have one anyway, and ask the frontier AI models to rank them. I think it would do a decent job of it. Of course, maybe it ranks differently than I do because it’s smarter than me. Maybe it gets the ranking right and I get it wrong. There’s always a dual hypothesis problem: if AI gets a different answer than me, is that because AI doesn’t have taste, or because I have poor taste?

Danny: I think, by the way, this mechanism is the reason why, subjectively, the rate of advancement has slowed down. For many problems I have in my workday, I am clearly not maxing out the capabilities of the model, and so I’m not feeling the frontier, because very few of the issues I have are actually complex enough to push it to its limit.

Bouke: So this brings up a really interesting question. This is something I’ve been discussing a lot with Philip at the AI Objectives Institute —

Danny: Philip Tomei.

Bouke: Philip Tomei, yes. Which is: what is the mapping from capability to usefulness? For my coding tasks, I’ve maxed out — it can do it. It doesn’t matter whether I use Opus or Fable. Sure, Fable is much better at it, but it doesn’t matter, because both solve it. That additional bit of capability gets me basically nothing. And I think we’re reaching that. It’s essentially diminishing returns to capability.

And that tells us a lot, because it informs us about how important something like recursive self-improvement is, and also how economically feasible recursive self-improvement is. If we are at an extremely flat part of the capability curve — which, by the way, isn’t true in some dimensions; I still can’t ask an LLM to solve cancer, it can’t do it, and if we get more capabilities maybe we can, so clearly there we’re not at a flat part yet — but for many tasks that people have, we’re pretty flat.

Danny: Flat in terms of marginal returns to additional intelligence.

Bouke: Exactly. In the sense that it can just do the thing we want it to do, and it’s very hard to come up with things that it can’t do right now, let alone if it were ten or a hundred times as intelligent. So insofar as that’s true, the returns to a lot of extra intelligence or capability in many domains are actually not that high — which means the economic returns to something like recursive self-improvement might not be that high in those domains.

Danny: But that’s assuming — and this is something I’ve been curious about and don’t have a good answer to — you can imagine you take a bunch of tasks and sort them by difficulty, or basically by marginal return to intelligence. There’s a certain segment where we get zero, flat returns to additional intelligence, and then some where that’s not true yet, and clearly that boundary is shifting: more and more tasks get eaten up by this boundary shifting. Two things I’m not clear on. One, where does the value accrue? Because you could still think the value accrues to those in the segment where the returns are indeed positive. And two, is this a partial equilibrium or not? We’ll find other things to do that are harder.

Bouke: A hundred percent. This is always the tricky part, because predicting what jobs will look like in the future is just so difficult. And I think there’s something to learn about that psychologically. That’s why we’ve had these predictions of mass unemployment for two centuries now: it is just so unbelievably hard to think about what jobs will look like if our current jobs disappear, or if machines can do our current jobs. And yet this has always happened, and every single time people failed at this. So it’s perfectly logical that we also fail at this. Maybe this time is different, but maybe it isn’t. Maybe we will all still be doing something that the machines either can’t do, or that we don’t want them to do.

And then there’s always a question of comparative advantage. Even if someone is better than you at everything, that doesn’t mean you’re out of a job. Acemoglu is arguably a better economist than me in all domains, and yet I write papers — it’s not like he writes my papers as well and I just write them worse. We just write different papers, because we have different comparative advantages. And that’s going to be true with humans and AI as well, as long as AI is expensive.

Danny: Exactly — as long as we’re not in a world of infinite resource in finite time, where there is no constraint on AI usage.

Bouke: Exactly. And at the moment, if there’s anything scaling laws teach us, it’s that getting more intelligence is expensive. Of course, you could get some kind of recursive self-improvement that would make the algorithm so much more efficient, or that would come up with basically infinite free energy or something like that, and then these arguments go out of the window. But we’re clearly not there yet. We are at a stage where, for some software tasks, AI is actually a really, really expensive solution. And that’s going to be true for several other tasks, where it’s just cheaper to hire a human — not because the human is better, but because the human is cheaper. We get specialization: we use the AI for the things it is comparatively good at, and we use humans for the things they’re comparatively good at, even if the AI is better at everything.

Now, on value accrual, I’m going to give a broader answer here, because this is something people don’t think enough about. People are hugely worried that the labs are going to make infinite amounts of money, and that this will lead to a society where a few people are immensely rich and everyone is poor. It’s really not clear to me that that’s going to happen, and that has to do with competition. There’s a reason that the current providers of electricity aren’t getting immense returns, despite their absolute prevalence in every single thing we ever do: there’s competition in the market for electricity. In the same way, it’s not like water producers can get infinite returns, even though we’d quite die without them. Competition means that producers, even if they create a lot of value, might not be able to capture all of that value.

Danny: And in these markets we’ve made sure, by regulatory means, that these are commodities or utilities that are distributed as such. We don’t let people extract ridiculous rents, because we make sure that there is competition, or quasi-competition.

Bouke: Yeah, in some ways. Quasi-competition. Most of these are actually state monopolies, for a whole different set of reasons; they haven’t worked very well as competitive markets. Although, I guess, electricity is quite a competitive market, whereas water is not.

Danny: Yes — I was thinking of electricity. It’s clearly not true of water.

Bouke: A hundred percent. Policy matters for whether you get a competitive market, as well as market characteristics. There is a reason Google search became a monopoly, and the main reason is that Google search became better because people used Google search. That gives rise to a dynamic where, even if your algorithm is not really that much better, you’re the first entrant, people use your service, your product gets better, more people use your service, your product gets better — which makes it entirely impossible to upend that market. Social media has a very similar dynamic. It’s why in the past twenty years we’ve seen huge concentration of tech firms that all became de facto monopolies.

For frontier AI, it’s really not that clear to me that this is happening. Open-source models that are 3% of the cost are four months behind the capabilities frontier. As long as that’s true, we’re not going to get massive value capture. At the moment I pay £200 a month for a frontier AI license. My willingness to pay for that is at least five times that. The moment my current provider doubles the price, I’m not switching. There’s a reason they can’t charge me a thousand pounds, and that is because there are competitors that provide a similar product. That competition brings the price, in the limit, towards marginal cost — which means these firms are not going to make huge profits.

Danny: That argument is, though, extremely sensitive to how economically meaningful the delta is between the leading edge and the trailing edge. And that to me seems to be one of the key questions, and I’m not quite sure how best to think about it: there is the gap as measured in benchmarks, but the much harder question is how economically meaningful that delta is.

Bouke: A hundred percent. I think this is a question we should be devoting a lot more resources to, which is: what’s the mapping from capability to usefulness? And that’s going to depend hugely on the domain, and it’s going to depend on the person, and we just don’t know. We have these METR graphs, and we have the Epoch graphs, and I think they’re incredible and super useful, but they don’t tell us this. They don’t tell us how useful that increase in capabilities actually is. So this is something Philip and I are trying to look into a little bit, but this is still early stage. And I welcome more people to join that pursuit.

Danny: What’s your current best stab at it?

Bouke: The current best stab would be something along the following lines — and obviously I’m revealing my alpha here. We know the capability frontier; we can take that from these benchmarks. We know the prices of models, so we sort of know the price of capability. And we have some idea about the demand, because people choose between these different models at different costs and at different capabilities, which tells us something about the revealed preference for how valuable an additional unit of capability is. I think combining those would give you, let’s say, the price of usefulness. The first maps capability and prices, and then from actual demand data you get some measure of usefulness based on people’s actual decisions.

And they might make mistakes — there are firms that have Copilot. Clearly not everyone is making rational decisions. Sorry, friends at Microsoft. But it is, I think, our best way of measuring this at the moment, because I think it is really important to have some kind of revealed preference over capabilities. Even if a lot of that is trial and error, we don’t actually know how useful that additional unit of capability is. But that would be my first step. What is your view?

Danny: One interesting test case that differentiates the AI-as-normal-technology view from the AI-as-maybe-not-normal-technology view is prediction markets. Metaculus runs competitions, and at present the very best humans are still just a smidge better than AI models. That is probably not going to hold for very much longer — maybe a couple more months. So it’s an interesting time: we’re about to see whether AI systems get better than the very best human forecasters. But I think what’s much more interesting is what’s going to happen a couple of months later, which is: is this where these systems top out? Because it turns out that the fabric of reality is such that you just cannot generate much more information out of what is out there, and so the returns to increased intelligence go to zero. Or do they just keep going?

Bouke: I think these are open questions in the capability-improvement literature. What’s the mapping from data to capability — and that’s both training data and context data? And are we going to hit a ceiling there? Can we have a self-improving system with purely AI-generated data? It’s not clear. A lot of computer science people are arguing that the model will just collapse at some point, because it keeps training on itself and doesn’t actually produce anything new, and therefore that’s the end of it. That’s why some people have said there’s clearly going to be a limit to how good LLMs can become.

I’m not a computer scientist, I’m not a model developer, so I’m probably the wrong person to ask this. But I think it matters a lot as a question: how much better can you get once you’ve surpassed the best of your training data? Can AI self-improve from there? And we don’t know. The current form of self-improvement that we have is just that we have AI write code, or better algorithms. It’s a bit unclear to me at this point what the mapping is from AI-generated data to increased capabilities, especially once you’re past the frontier. And I think that matters. It will be a huge determinant of how big a deal recursive self-improvement is going to be.

Danny: If we go back to your published research versus the as-yet-unpublished things: you created a really interesting exposure measure that measures how amenable to reinforcement learning different jobs are. So rather than measuring, backward-looking, which proportion of the tasks that make up a job we think an LLM can do at a reasonable level of quality, what your index measures is whether we think, with our current understanding of reinforcement learning, that we can train a model to do the tasks that then make up this job. And of course the interesting question is where the two views diverge. So what are some examples where we would think a job might have low exposure as measured by some of the classical exposure measures — Eloundou and others — but high RL-ability, or high amenability to being trained against with reinforcement learning?

Bouke: That’s a great question. What Eloundou et al. measure is: you have all the tasks, can these be done by an AI? Whereas what we measure is, for this particular task, can you create a reinforcement learning environment where an AI learns to do the task better and better? And as it turns out, these do differ. One clear job is CEO. If you think about what jobs are RL-able, it is stuff where the rewards are very valuable, the outcomes are clear, the actions are clear, the action space is clear, and the mapping from actions to outcomes is clear. There are more factors, but I think these are the big ones. For the CEO, no — because there is an infinite action space, and it’s super unclear how the actions map to outcomes. That’s why measuring CEO quality is such a hard job. That’s why it’s just not an RL-able job: most actions are taken once, so you can’t learn anything from there, and you don’t actually know what outcomes those actions caused, so we can’t learn from there either. Whereas if you just look at the task descriptions, anything that is text-based is AI-doable. CEOs were actually quite high exposure in the index; they’re really not high exposure in terms of our RL index.

And then sort of the opposite: jobs where there’s not that much text-based stuff, but where you have a limited set of actions, those actions cause very clear outcomes, and the whole space is very clear — operators, for instance. Let’s say you operate a plant. What that mostly means is you press the right buttons at the right time. You know what the actions are, namely the buttons. You know what outcomes those buttons cause, because most of these plants measure everything in real time, so you know exactly: button is pressed, x happens. That’s super RL-able, because you could literally simulate the entire environment, learn what actions lead to what outcomes, and just do the job much better than any operator ever would. And this is a super low Eloundou-et-al. job, because it doesn’t have text-based stuff. So these are, I think, the two groups where our measure really diverges from theirs.

And if we think about it, I actually think this RL stuff is really quite important, because this is what the labs are doing. The labs are actually trying to create these reinforcement learning environments for many jobs, where they effectively try to film people doing their job and see whether they can turn this into an RL environment — that is, can we create the data that allows us to learn this job? This is why Meta is documenting everything their employees are doing: basically, they try to RL their employees away. Now, for obvious reasons, people are quite unhappy with that. And many startups are doing the same thing: they’re trying to create RL environments of jobs so that agents can do that particular job. So I think this will tell us quite a bit about where future improvements are going to come from.

Danny: Luis Garicano has this notion of messy jobs — the amount of messiness — with the argument, of course, being that the messier a job is, the less amenable it is to automation with LLMs. What is the mapping from RL-ability to messiness?

Bouke: Yeah, I think this is a great question. Because our measure is still a task-level exposure measure: can you learn this task? We’ve also done it at the job level: could you learn the whole job through a reinforcement learning environment? But it’s always a bit of a question of exactly what that means. So our measure tells us which tasks can be learned by AI. Whether that task actually disappears from a job bundle — even if it can be done by AI and there are no regulatory barriers — depends on why it is part of the bundle to begin with. Can you unbundle this job? And that’s where I think Luis’s work is great. I think the idea of weak and strong bundles is a very important one.

If I think about my own job and I were to decompose it — I do research, I teach, I make exams, I grade exams, stuff like that — it’s not clear to me that I could unbundle making exams. AI can make great exams, probably much better exams than I do. But me making and grading the exam helps me to teach better. And insofar as you have these complementarities and these spillovers between tasks, even if AI can do a certain task, for my employer it would actually be rational to say: no, you keep on doing the marking and the exam-making, because you become better at the task we can’t automate, namely the teaching. And if you have those spillovers, that would be a reason to have a messy job — it’s a variation of a messy job. Even if AI can do it, the firm might not want to automate that task.

There are other reasons that jobs are messy. It can be that it’s really unclear — you have five tasks, and you don’t know how each of these individual tasks maps to an outcome, but you do know how the composite maps to an outcome. I can decompose my research into different tasks, but I don’t know which one of these is actually driving my paper quality. Therefore, if you want to reward me based on paper quality, you have to keep the tasks bundled. That’s another reason jobs aren’t unbundled. So I think the weak-versus-strong-bundles paradigm of Luis’s is super useful to think about when automatability actually leads to automation, even if there are no regulatory barriers, no adoption barriers, et cetera. So I think they’re super complementary, in the sense that we need both these elements to understand what jobs will look like in a few years.

Danny: Right. If RL-ability gives you the conceptual floor — can we do this in the first place? — then the messiness logic gives you the answer to: well, even if we can do it, would we want to? Will it happen?

Bouke: Exactly. Is it in the self-interest of a firm, or an employer, or an employee for that matter, to do this? Because it might indeed then reduce productivity on other tasks. It’s a little bit like the weakest-link logic.

Danny: Right — the O-ring logic.

Bouke: Yeah, exactly. The O-ring logic is a very, very strong example of that, where if you have complementarities between tasks, it might actually make sense to keep some, especially if doing one task makes you better at another task. I think Luis’s work is great. I see it as complementary, and this is something Jakob Schell and I are working on a little bit. We actually know quite little about why jobs are bundled together. We have a few theories, but this is not an extremely mature literature. It’s been going on for seventy years: in the classic Adam Smith pin factory, the whole point is that if you do one task, you just become so much better at this, so we should all specialize. Why aren’t we specializing? Well, because there are other coordination frictions between tasks.

So there’s a great paper on sequential tasks by Peyman Shahidi, John Horton and some others. It matters a lot, if you have four tasks in a row, whether it’s tasks one and three that are automatable or tasks two and three. Because every time you delegate a task to the AI there’s a communication friction — you have to communicate it to the AI, and the AI has to communicate it back to you. Whereas if there are two automatable tasks in a row, that friction only happens once. So the sequence actually matters quite a bit. It matters not only how many tasks, or which task; it matters where in the sequence of tasks.

Danny: You would expect, similarly, that the length of the value chain matters a great deal. For things that have a very short value chain, if they’re automatable, you would expect that to happen right away. Whereas for things that have a longer value chain, even if, to your point, individual steps within it are automatable, you would expect that to take much, much longer. Which weirdly connects to a question that sometimes comes up: why can some firms offer a four-day work week, say, and others can’t? And I think it has everything to do with how long your value chain is, and how many handoffs there are. If the answer is that there are actually very few handoffs, that’s very easy to dial up and down how much someone works. Whereas if there are lots of handoffs, then it’s much, much, much harder, for the exact same reason.

Bouke: I wholly agree. I mean, okay — I’m Dutch. The entire country works four days a week. I think the average Dutch person works something like thirty hours a week. Clearly, it seems to be going okay. I mean, people at ASML have forty holiday days a year. That’s pretty good, and it hasn’t stopped them from making arguably the most complex machine in the world. And partly, you know, they have so much money that they can give their employees anything. So maybe it’s that.

Danny: Maybe it tells you more about the value of the individual employee.

Bouke: Yeah, that’s probably true. So the question is sometimes: if you double your number of employees and they each work twice as little, and keep capabilities and everything the same, does it work better or worse? And that depends on what the diminishing returns to time are. If I double the amount of time I work, I’m not going to double the amount of work I do, in the sense that I just get tired, and I run out of things I can do productively. If I work six hours a day, I’m going to be more productive on average than if I work eight or ten or twelve hours a day. My total productivity still increases, but my average productivity goes down. And at the same time, if you have a lot of people, there is a lot more coordination cost and management cost. So which of these two is stronger determines whether you can do four-day work weeks with more people, or whether you need a smaller team that works a lot.

That’s the private equity model. Basically all private equity firms, as far as I’m concerned, are really quite small, and everyone works eighty to a hundred hours a week. Surely that’s unpleasant for them. If it were easy for them to just do the same job with twice as many people, they would probably do it.

Danny: Right, but they can’t. So I imagine there’s a reason that they’re not doing it, and it’s got to be something about coordination costs.

Bouke: Yes. I think consulting firms are similar: small teams, very high work hours, and I think it has to do with the cost of handovers.

Danny: The friction of handovers, and the importance of context. The returns to having a lot of context are very high, and that’s hard to share across brains.

Bouke: Exactly. And that’s going to be a big determinant of which tasks you would want to automate. I think another really interesting point there is outsourcing. If we want to think about which tasks we can unbundle, look at outsourcing. I don’t know the outsourcing literature super well, but if I were to study which tasks we can unbundle, that’s what I’d be looking at. Clearly these are tasks that we could take out of the firm and do elsewhere, and clearly they’re tasks that in some way we can verify. So in a sense they might be both RL-able and outsourceable, and unbundlable. And therefore these are the tasks that we can let AI do — because we can verify it, we can RL it — and it is unbundlable, so we can let AI do it rather than let someone in another country do it.

That’s where I expect a lot of the unbundlable tasks are. And that’s why I’m actually really quite worried about countries that are current recipients of outsourcing. And, again, I’m not a development economist nor a trade economist, but I think we should be worried about this particular issue: countries that are currently developing through being the recipients of outsourcing — that might disappear, or at least go down a lot. And insofar as that’s true, these countries need to come up with a different development model. And as it turns out, that’s quite hard. Maybe we can ask AI in two years.

Danny: It’s entirely non-obvious what that replacement looks like. There’s some work out there, but there’s a surprising dearth of work on this.

Bouke: Yeah. My theory on that is that the incentives for development economists — I mean, this is a Twitter debate every year — are to run RCTs on relatively small questions, relatively doable questions, rather than on macro development in some way. Why are some countries rich and some countries poor? What are the determinants of structural change? Why is India doing services rather than industry? As a result of that, insofar as development economists have largely specialized in running RCTs, they just don’t have a comparative advantage to think about how AI will impact a country. That’s not an answer to your question.

Danny: Lant Pritchett has been making this point for a long time, and I think he’s right.

Bouke: Yeah, no, I think so. And this is, to be fair, a debate that has been coming up in development economics many times, not in the context of AI. But it is, I think, a potential reason why too few economists are thinking about this problem, and we should get more. The effect of AI on the developing world is clearly understudied.

Danny: By a wild, wild, wild margin, given the importance. If you have an export-led development model, particularly for services like business process outsourcing, but not exclusively — potentially also for manufactured goods — it’s not obvious that any of this is good news. And yes, some of the answers you may have: the deflationary effects of AI may be great, you may be able to provision services more cheaply, it may be great for your health care system or the rest of it. It’s not to say that it’s purely bad news. But in terms of economic development, it’s not clear what that ladder looks like if AI capabilities keep improving.

Bouke: No, a hundred percent. And of course you can make sort of the opposite argument: if you have a phone and you have internet and you live in a remote part of India where schools are bad or the teachers don’t show up for class, suddenly you have the most incredible opportunity to learn everything. It’s not clear to me that people are worse off, whether in remote India or in London.

Danny: But this is the MOOC thing. We’ve had MOOCs for a long time, and their effect seems to have been negligible.

Bouke: Which is — I would not have predicted that. So I will be humble here in my predictions about the future of education. It’s in a sense quite incredible that even five years ago you could get a much better economics education than anyone could ever provide you with, because the absolute very best economics teachers would just record their lectures, put them online, put their exams online. You could learn everything from them, let alone the entire universe of YouTube tutorials and stuff. Nonetheless, the demand for education has only gone up — for further education, especially at good universities. So clearly that didn’t pan out. And I’ve tried to do MOOCs; I can’t motivate myself for it, and I’m a pretty high-intrinsic-motivation kind of guy. So I totally get it. Universities provide a lot more than just the ability to learn, or the lecture slides. We provide the incentives. We provide in-person signaling. We provide network effects. We provide all these other things.

Danny: Dating markets.

Bouke: Dating markets, absolutely. I met my wife when I was at university.

Danny: There you go.

Bouke: Different university, but that’s slightly beside the point. Especially now that people are souring a little bit on dating apps — or maybe that’s just my environment. But clearly, providing these opportunities for people to meet in person, in a world where people meet increasingly little in person, I think is super valuable.

And — sorry, I’m going to take a short detour here, maybe we’ll come back to what we’re actually talking about — this is what I sort of expect I might be doing if AI fully automates my research. This is a point both Alex Imas and I have made at some point or other: humanness preferences matter, in the sense that for some things a human being part of the product is part of the value, even if it is worse in some kind of objective quality. The obvious examples here are that we’re not going to watch robots do ballet, and we’re not going to watch robots play the World Cup. Because the fact that it is humans is part of the thing that makes it interesting to us.

Danny: Anton Korinek has called this nostalgic jobs.

Bouke: Nostalgic jobs. I would actually strongly disagree with that. I think there are also nostalgic jobs — we have preferences for handmade stuff, and I think that’s a little closer to the nostalgia point.

Danny: Fair enough.

Bouke: But I think it’s much more a much deeper intrinsic preference we have for interacting with humans. So I could easily imagine that what I’ll be doing is running highbrow reading groups, where the AI has got incredible —

Danny: I’d turn up for that. That sounds amazing. I’m here for it.

Bouke: But basically, the AI is going to do all the research and come up with all these incredible new insights about economics. There are a whole bunch of people who are interested in understanding economics for no instrumental reason, except for the fact that they’re interested. I have the expertise to discuss it with them. At the moment, we have so many book clubs. Is that because we can’t find a good online discussion of the book? No, obviously not. We want to discuss the book in person with people, because we enjoy that in-person interaction. That is the value proposition. And I think that’s going to be true in many domains. And I don’t think it’s fully implausible that that’s indeed going to be what most of us will be doing — we’ll just be entertaining each other in some way, for things where that humanness matters.

Danny: I think this is a great example, because I think often this argument gets made in what I think is a lazy way, which is to say: well, we have preferences for being served by humans in certain interactions. Which I think is true — that’s what you’re saying. And then often the argument is, well, for instance, take medical services. And that to me seems like a supremely bad example. It’s not obvious at all that people on average will prefer speaking to a human doctor. And in fact, relatively soon, if not already now, I would be worried in some sense if I had a complex medical problem and my health care provider was not using LLMs to augment what they know. And it’s not obvious that speaking to the human there will give me anything at all, maybe barring some weird edge cases. So I think your example is a much more plausible one.

Bouke: I wholly agree that this is a super bad example. I actually think I probably derive negative value from interacting with humans in some cases. Imagine I have an embarrassing medical problem, or imagine I’m in a nursing home and I need someone to wipe me or feed me. I find it really embarrassing to have to do that with a human. I’d much rather have a robot do it. There are other things where that’s not true — there are parts of the medical system where I like to discuss with a person, because I value that someone reassures me. I think there’s some emotional support that doctors give that is valuable. So I don’t want to fully dismiss that humanness matters. But there are many things where that’s just clearly not true.

Danny: It’s true. But we also have great studies on a lot of doctors having absolutely terrible bedside manner. So I think sometimes there’s an idealized version of this that gets bandied around, which just does not track with reality.

Bouke: Well, that would just change the job description of doctors.

Danny: Exactly. It would make medical school much less important.

Bouke: Maybe this is where the future is going to go: rather than training people to code and to learn economics, we’re just going to be training people to be good entertainers, or better humans in some way. What an incredible world that would be — where, instead of training people to learn Gray’s Anatomy by heart, we teach them bedside manners, and how to reassure people, and how to make people feel at ease. In my case, how to interact with people so that you get a great discussion going. I mean, this is obviously utopian. This is not a prediction about where we’re going to go, but I would love to see that world.

Danny: I’d be down with that for sure. Another big question, which relates to this, that we don’t know a ton about, is the elasticity of demand for different jobs. If we want to know not just what the exposure is, but what the net effect is going to be, everything turns on that. How should we think about this?

Bouke: A hundred percent. I wrote about this in January this year — or last year, I forget. Because basically I was writing three blogs on a positive vision on AI and labor, and I never published any of them. But I sent one of the blogs to Sarah O’Connor at the FT, and we sort of did an interview exactly about that point. And then Alex later on wrote a very similar blog that was arguably much better written than whatever blog I was working on. So I’m very much on board with that.

But if you go back to the 1920s, we got mechanization in agriculture and in the car industry. Why did the number of people working in agriculture go down five times, and in the car industry go up by five times? In both cases the same amount of stuff was mechanized. Prices in both industries went down by similar amounts. What explains it? Well, it’s the elasticity of demand. Halving the price of food means we buy 3% more food. That just means that if you automate away half of agriculture, we’re just going to have half as many farmers. For cars, a 50% or 20% or whatever decrease in prices meant demand went up ten times. So yes, one person produces twice as many cars — but we need ten times as many cars, so we need five times as many people working in the car industry.

This is a hugely important question, and we have such poor estimates of elasticity of demand at the occupation level. OpenAI tried to do this, where our best estimate is OpenAI asking ChatGPT what it thinks the elasticity of demand is. That is, for me, a super unconvincing measure.

Danny: That seems suboptimal.

Bouke: Yeah, that is clearly suboptimal. And, I mean, to be fair, economists have been trying to estimate elasticities of demand for some products; it’s not that we have nothing. But to do this at the occupation level — we have no idea what the elasticity of demand for consultants is, or for software. We have maybe some idea about the elasticity of demand for shoes. I’ve actually written a paper on the elasticity of demand for medicine in a very specific context, but that’s beside the point. We have some estimates, but there are so many jobs where we just have no idea. And I’ve written some proposals about how we can do something here, but it’s really a first step.

Danny: How could we measure it better?

Bouke: Anything where we have high-frequency consumption data, we can get somewhere. Let’s say there are these scanner datasets where, across supermarkets, you know what the prices of things are, and we can match that with the employment distribution of the products that firms are selling. So we know that whatever firm is selling product x in a supermarket or in a shop, the employment distribution of the firm is 10% software engineers, 30% x y z, et cetera; labor is 40% of their total costs; and this is a super competitive market. If we have all these elements, and then we get a cost shock — which happens all the time; it could be an interest rate shock, it could be a change in the price of the inputs — that changes the price of the product without directly affecting demand. That’s some exogenous —

Danny: Some exogenously induced variation in the price.

Bouke: Exactly. So we have lots of instruments for that. We see a change in price: how does that affect demand? And if we know who’s producing that, we know how that changes the demand for those particular professions.

So: prices go down by 30%, because we automate 30% of the job — because this is the thought experiment we want to have. And we know the cost link, because it’s a competitive market, and therefore it will have to get passed on to consumers, which is why we’ll see it in prices. So we need to know that link too. Imagine we automate 30% of labor. To then understand how much that changes costs, we need to know how big a part of costs labor is. So if it’s only 10%, you can automate all labor away and the prices are not going to go down very much. So that link tells us how much costs can go down. Then, to understand how costs translate to prices, we need market structure: if it’s a competitive market, a change in marginal cost will lead to an equivalent change in prices. And then the elasticity of demand tells us how that change in prices affects demand — which then tells us: okay, now we’ve automated 20% of the job, but we have two times the demand, so now we actually need a lot more workers. So these are all the different links that we need to know. And the main open question there is the elasticity of demand for products. We can do that with scanner data, but that’s a limited set of products; it doesn’t tell us about the consultants, the software engineers. So we need a lot more than that, but this would be my first step.

Danny: It stands to reason that the difference — just like in the example with the car industry and agriculture — will be huge. One of the obvious examples is bookkeeping. I would assume, and I could be wrong, that the amount of bookkeeping I’m going to consume as a corporation is more or less fixed.

Bouke: I think so. I sort of expect that to be true. There will clearly be jobs with low elasticity of demand. Bookkeeping could be one, although there are bookkeeping-adjacent services, like some kind of financial analysis, where you might just —

Danny: Maybe we do more controlling.

Bouke: Yes, who knows. Exactly. There are so many things we don’t check at the moment, where you might check every single invoice to see that it is correct. So it’s not clear that it is fully constant, but I sort of agree that I don’t expect that to be a super elastic product.

At the consumer side, what I think elastic products often are, are things that are currently luxury products. So the example I used with the FT was interior designers. I work on a UK academic salary; I can’t afford an interior designer every time I move. If it would cost 10% of what it currently does, I would get an interior designer every move, just to tell me which colour couch I should have and how I should structure my place in a way that makes sense. In the same way that I would get a full-body scan every month if that were really cheap. There are all kinds of things that are currently luxury products that I can’t afford, or that have clear value in the sense that they’re not just status objects. So basically, if I want to look at high-elasticity-of-demand products, I would look at current consumption patterns of rich people, taking status consumption out of the equation. Status consumption is also relational — that’s actually what Alex Imas wrote about. But I think there’s a lot more here that we can learn from rich people’s consumption.

Danny: Another thing you’ve written about is the apprenticeship pipeline problem. We’ve talked about this earlier. Presumably the trade used to be: I, as a corporation, would take on someone relatively junior; they wouldn’t provide amazing value, but I also wouldn’t pay them huge amounts of money, and so everyone ended up happy with the trade. Now I’m limited in how much less I can pay them, to a degree, but their usefulness goes down, and therefore I’m less likely to want to hire someone. The thing I don’t understand about this story is that over time — on the assumption that firms will want to continue to exist, and on the assumption that you somehow have to train people somewhere — won’t they have an incentive to fix this?

Bouke: So Gary Becker wrote about this in the sixties. Basically, this problem exists independent of AI. And his basic point was that you can’t contract future productivity improvements. So I can train someone, but then someone might just poach them.

Danny: This is why firms underinvest in training.

Bouke: Exactly. This is why firms underinvest in training, and the value proposition has always been: okay, we’ll invest in some training, or on-the-job training, but we’re going to pay you less than you actually deserve. And as far as that goes, it’s really not clear. Because if you have two firms — those that train and those that don’t train — the ones that don’t train have lower costs, because they don’t spend those resources training. So the moment that person becomes trained, they just poach them.

Danny: Isn’t the answer, though, kind of obvious? Which is the same thing we do with training: we subsidize it. And we could just subsidize junior employees.

Bouke: So you could subsidize junior employees. You could subsidize training. You could try to make junior employees more productive.

Danny: How would you do that?

Bouke: Let’s say by — currently, if I have an incoming PhD student, I would want to train them on how to use AI. That would make them much more productive at the job they’re doing. So it’s either learning with AI or learning to use AI; I think these are the two obvious ways in which we can get there. Now, the last solution, which I’ve talked about a little bit with some people and everyone hates it, is that you could theoretically contract future productivity improvements.

Danny: In higher ed, that’s income share agreements, which do something similar.

Bouke: Exactly. But I think they’ve been ruled illegal, if I’m not mistaken. Because the point was — and I think it was correct — that there are people who have liquidity constraints and who are really smart, and therefore you have a market failure where no one is willing to provide them the loan for an education, even though they could pay it off from their future earnings. And to solve that market failure, people contract future earnings, saying: okay, I’m going to pay for your education, but you owe me 10% of your future earnings. That’s exactly the kind of system that I was going to talk about, where you could say: anyone who employs people in the first five years of their career becomes eligible to get x percent of their future earnings.

Danny: It’s elegant, but I can see why people would hate it.

Bouke: Yes. I think it solves a problem in two ways. First, it just makes it more attractive to hire a genius, because you get a future return. And second, if you invest in them becoming more productive while they’re at your firm, you gain. Of course, you could say, well, maybe the government would then just pay for that amount, which makes it slightly more palatable to people, but less to the government. But I think these are broadly the solutions: you have to make people cheaper, or more productive, or you have to contract capability improvements.

Danny: The flip side to all of this — and this is going to be my last question — is this idea of pro-worker AI. So we can think about how to fix the worker, as it were, or we can think about how to fix the state of the technology. Probably it’s going to be both. I understand the notion, of course, but I’m struggling to understand what that would look like in actuality.

Bouke: I think there are two things here. We’ve all seen the AI statement of — what was it, yesterday, two days ago? And one of the sentences in there is indeed about pro-worker AI in some way or other; I forget the exact wording. And there are obvious cases where automation is good. I think having a world of self-driving cars, despite all the labor displacement that will cause, is a better world. Or having teachers in remote parts of the world, where currently there are very few teachers, that are fully AI — probably an improvement. So it’s not clear to me that we should only have pro-worker AI.

Now, on top of that, it’s a little bit of a question what that means, because there are so many emerging capabilities that we don’t understand how they came about. At the moment I think my AI is super pro-me, in the sense that it’s just made me hugely more productive in terms of what I do. Has anyone designed it that way? Absolutely not. These are just emerging capabilities from a model that has grown rather than been designed. I actually quite like the Anthropic framing here, that we don’t design these capabilities, they emerge. Sure, some are post-trained, and there’s something we can do there. But post-training, making stuff more pro-augmentation and less pro-automation — I’m skeptical that we can do it, and I’m skeptical that we even know what that looks like. It’s not clear to me which dimensions we would tease out from a model versus what we would suppress. Practically, I lack the understanding of what that would mean.

Let’s take one task — let’s say coding. What on earth is an augmenting coding agent compared to an automating coding agent? I just can’t grasp that. I don’t write code anymore. I still work with AI to write the code. Has my coding been automated in some way? Do I still do coding stuff? For now. I don’t even know if that means it is augmented or automated, because that augments all my other stuff: by automating coding, I can do much more other research, I can explore ideas much quicker. It augments me even if it automates the task — and there is a software engineer somewhere else who only did the coding, and they’re fully automated. Should we have that? Should we not have that? Any given task is augmented for some and automated for others. I just don’t think this is feasible, nor realistic.

Danny: With that, on that happy note — Bouke, thank you very much. This was a great deal of fun.

Bouke: Entirely my pleasure.

Danny: Thanks for listening to High Variance. You can subscribe to this podcast on Apple Podcasts, Spotify, or wherever you get your podcasts. If you like this podcast, please give us a rating and leave a review. This makes a big difference, particularly for newer podcasts like this one.