Worth doing, even if a machine does it better

Whenever I show something I’ve made with AI, I get some version of the same question. Under a post about Graviton Theory, my music project, someone joked that the next step is to have the AI generate the artist as well, to get rid of that pesky human interference. After a talk I gave on AI for scientific discovery and automated research pipelines, a professor told me that research a machine does end to end, from the idea to the finished paper, with no human involved, is worthless.

Both have a fair point, and I don’t want to wave it away. But I’ve had this conversation often enough that I want to write down where I’ve landed, so I can answer with more than a comment.

My view is that three different questions get mixed together whenever this comes up:

1. Function: what has to get done for a society to work and for living standards to keep improving.
2. Value: what a person finds worth doing, which each of us decides for ourselves.
3. Frontier: what we’re able to do at all, and what we find out by trying.

Automation changes the first question a lot. I think it shouldn’t change the second, and it opens up the third. Much of the worry I hear comes from the fact that for a long time the first two were tied together: you were valuable because of the job you did.

Work a society needs done

Some work has to happen for a society to function, if we assume that surviving and living better is the goal and that we all want to benefit. Food gets grown and delivered, power plants get maintained, accounts get balanced, taxes get filed, software gets tested. Some people love parts of this work, others would never choose it, and much of it is somewhere in between. I don’t want to rank these jobs. Bookkeeping can be the fulfilment of one person’s working life and a chore for the next person.

From this angle, automation is a practical question: can the work be done reliably, faster or cheaper, and who’s accountable when it goes wrong? It’s the question we asked about looms, tractors and spreadsheets. John Maynard Keynes asked it in 1930, in Economic Possibilities for our Grandchildren. He expected that “the economic problem may be solved, or be at least within sight of solution, within a hundred years”, and that people would then face “his real, his permanent problem”: “how to use his freedom from pressing economic cares, how to occupy the leisure […] to live wisely and agreeably and well.” We’re four years from his deadline and the economic problem isn’t solved for most people. Still, I think he named the second question correctly.

Value is something people assign

The second question is about value, and the answer is easy to state and hard to live by: something is valuable because a person values it. For one person, an evening at a concert is the best part of the month. For another, watching twenty-two people chase a ball and try to get it into a net is a waste of time, while for millions of others it’s the highlight of the weekend. We accept this without discussion for hobbies, sports and art.

The idea is old. Carl Menger, who founded the Austrian school of economics, wrote in his Principles of Economics (1871): “Value is thus nothing inherent in goods, no property of them […] Hence value does not exist outside the consciousness of men.” Nietzsche put it more dramatically in Thus Spake Zarathustra: “Through valuation only is there value.”

What’s new is that we now have to apply this to work as well. Most of us were born into a system where the job you do is how you’re valued. You specialize, you become useful to the system, and that usefulness turns into how you value yourself. Contracts, salaries, titles and the “so what do you do?” at every dinner reinforce it. I think this link is the problem, and it’s what scares many people right now. If a machine can do the function, and your worth came from the function, your worth seems to go with it. Knowledge work is where the two are coming apart first.

I’d argue the link was always a convention. As a kid I never dreamed of doing my own tax returns. Today bookkeeping, accounting and admin take a good part of my week, because they come with doing research, experiments and art. I didn’t sign up for that as a child, but I accepted it as the price for the rest. If machines take over parts of it, I lose a chore and keep everything I care about. Someone else may love exactly that work, and then it’s worth doing for them, whatever a machine can do.

Learning works the same way. People pursue knowledge because it changes them, and that has value of its own, whether or not a machine can produce the same result faster, cheaper or a thousand times over.

We still play chess and Go

Games are a good test of whether a human activity loses its value once machines are better at it. Deep Blue beat Garry Kasparov in 1997, and AlphaGo beat Lee Sedol in 2016. Today no human has a realistic chance against the best engines in either game. People kept playing anyway, and on at least one platform in growing numbers: Chess.com reported 100 million members in December 2022 and 250 million in February 2026.

Go shows both sides. Lee Sedol retired in 2019 and told Yonhap: “Even if I become the number one, there is an entity that cannot be defeated.” For him, being the best was part of the meaning, and I take that seriously. At the same time, a 2023 study in PNAS analysed 5.8 million moves by professional Go players from 1950 to 2021 and found that “humans began to make significantly better decisions following the advent of superhuman AI”, with more moves that had never been played before. People learned from the machines and kept playing, because they value playing.

The comparison has a limit, and it’s the important one. Human play and learning kept their value even as engines got stronger, but that doesn’t show that knowledge workers whose jobs get automated will keep their income.

Prices and wages are how a society turns many subjective judgments into something that looks objective. Menger’s point applies to markets as a whole: a price is what people together are willing to give for something, and a wage is that judgment applied to someone’s work. Hayek described the price system as a mechanism that passes on knowledge spread across many people. Contracts, labour law, taxes and social insurance are built on top of it, so in practice we credit people with value through what their work is paid. When a machine does that work for a fraction of the cost, its price falls, and under today’s rules the person’s income and standing fall with it, although nothing about the person has changed. Without a replacement, whole groups of people and professions lose their footing.

I see this as a shortcoming of how our societies are built today. Changing how we credit value to people, in law, in government and in society, is macroeconomic and political work, and it won’t sort itself out.

Everyone defends their own field

The professor’s argument went further, and this is where I disagree with him. In the same conversation he said he likes that AI now writes his unit tests: it saves him tedious work and, as he put it, has made him a much better programmer. He also said that the mathematicians who have to check the AI-generated Navier–Stokes proof should be paid something like 10,000 euros an hour. (I’m paraphrasing from memory.)

I agree that checking this kind of work is skilled work and should be paid well. The same holds for the testers, QA engineers and DevOps people who have checked software for decades, in research labs as much as anywhere. Unit tests are the programmer’s version of checking a proof. If a machine writes them and you skim them, you can’t tell whether they test the right things, the unknown unknowns stay unknown, and no one is paid to look for them anymore. I don’t see how that makes anyone a better programmer.

I don’t think the professor is unusual here, and I catch myself doing the same. People defend their own field and are fine with automation everywhere else: research is human and creative, so it shouldn’t be automated; testing is tedious, so it’s fine. The writer Joanna Maciejewska put the popular version in a 2024 post that went viral: “I want AI to do my laundry and dishes so that I can do art and writing, not for AI to do my art and writing so that I can do my laundry and dishes.” I understand the feeling, and nobody should have to do chores to earn the right to make art. But each of these lines splits work into the part that matters, which is ours, and the part that doesn’t, which belongs to someone else. Each of us can assign value to our own work. We can’t take it away from someone else’s.

The split between creative and functional work doesn’t convince me either. Coming up with an idea, writing code, writing a paper and designing a test all involve choices nobody hands you, and all of them can be creative. So can art. Calling research creative and testing mechanical reduces both.

Why automate research at all

Behind the professor’s remark is a fair question: why would we want research automated in the first place?

I have two answers. The first comes from systems theory: we already live by relying on systems we don’t understand in detail. Most of us never wrote our own operating system, designed and built our own laptop, or built the power plant behind the socket. We know the interfaces: the plug fits, the voltage is what the label says, and Kirchhoff’s laws tell us what happens in the circuit. That’s enough to use these systems for our own goals instead of spending our lives securing food and energy ourselves. Herbert Simon described an artifact in The Sciences of the Artificial as “a meeting point” between an “inner” environment, “the substance and organization of the artifact itself”, and an “outer” environment, “the surroundings in which it operates.” Alfred North Whitehead wrote in 1911: “Civilization advances by extending the number of important operations which we can perform without thinking about them.” If we understand the interfaces of automated research well enough, it becomes one more of those operations.

The second answer is complexity. Our working memory holds about four chunks of information at a time. Problems like cancer, diabetes, the energy transition or climate change have so many interacting variables that it’s hard even to decide which ones to look at first, and every intermediate problem on the way needs solving too. Systems that scale better than people can search more of that space, orchestrated by people and perhaps later by other systems. AlphaFold is the clearest example so far: according to the Nobel committee, Hassabis and Jumper “have been able to predict the structure of virtually all the 200 million proteins that researchers have identified.” The Navier–Stokes work is another, and it shows the human side as well: Quanta reports that both AI teams working on it relied heavily on an approach developed by Diego Córdoba and Luis Martínez-Zoroa.

None of this leaves people without a role. We build the machines and decide which problems they work on. We also decide what a result is worth and whether to act on it, because we’re the ones living on this planet: a treatment or a climate solution matters once people approve it, regulate it, pay for it and use it. We could run two systems side by side, one fully automated and one run by people, and both would have value, the same way engine chess and human tournaments exist side by side.

I went into how such research systems are built in my talk Agentic AI for Scientific Discovery at the RL Bootcamp 2026 in Salzburg (recording); the two arguments come up at 2:25:34.

The frontier is a reason of its own

The third question is about capability: trying things to see what’s possible. Some of it is a demonstration, some is curiosity, and some leads somewhere nobody planned. Graviton Theory started like that. I wanted to see what happens if AI is part of the whole creative process, from the music and sound to the visuals, videos, artwork and storytelling around an artist. Several albums later it has become a project of its own, and it let me make things I couldn’t have made alone a few years ago. The lyrics and production decisions are mine; the visualizers are 3D films rendered in Blender by a pipeline I built with coding agents.

Even a machine making art end to end, with nobody in the loop, can be worth doing as an experiment. It might teach us something, at the second or third remove, about creativity or discovery. It might be an ingredient for far more capable machines later, maybe even for questions about machine consciousness, though that part is speculation. Or it might turn out to be worth nothing to anyone else. I’d still count it as valuable, because finding out new things is something I value, and by the argument about value that’s enough.

Exploration like this also creates new questions for society and for individuals, and some of them are burdens. I don’t want to play that down.

What I’m claiming

When I share a view like this, people sometimes read it as a plan for how society should solve the whole problem, then point to the parts it doesn’t solve. They’re right that it doesn’t. I’m one piece of this, not the one steering it. I can’t speak for artists, and I don’t claim to be one in the traditional sense. I see myself as a music producer with my own story, thoughts and feelings, which I express through this process. I don’t play an instrument, and I don’t think that makes what I made less valuable. The same goes for research: if the pipelines I work on one day write better papers than I do, that says something about the pipelines and nothing about my worth.

Changing how prices and wages credit people is a question for everyone, not only technologists. It needs new laws, new guidelines and new institutions, and it needs discussions that include all the people affected. Capable machines will eventually do most work that can be done digitally, probably better than any person. If our current system only values people for that work, I don’t conclude that people lose their value. I conclude that the system needs to change.

What makes this harder

Work was never only a function, which makes separating it from value harder than it sounds. Hannah Arendt worried in The Human Condition (1958) that automation would leave us with “the prospect of a society of laborers without labor, that is, without the only activity left to them. Surely, nothing could be worse.” Marie Jahoda, whose 1933 study of the unemployed in the Austrian village of Marienthal is a classic of social research, later described what a job provides besides pay: a time structure, social contact, a collective purpose, status and regular activity.

I agree with both. Structure, contact, purpose and status matter, and they can come from more than economically necessary work: communities, clubs, research groups and art projects provide them too. We don’t know yet how to provide them for a whole society, and getting there could hurt many people on the way.

Money is the second complication. “Value is what you assign” doesn’t pay rent. Even if we agree that the way we price work has to change, that change will be slow, and until it comes, separating function from value is much easier for people in my position than for someone whose job gets automated next year.

Third, freed-up time doesn’t automatically go to the things people value. Keynes thought “three-hour shifts or a fifteen-hour week” might be enough work for most people, and we’re far from that. Time freed by machines can just as easily fill up with new kinds of busywork.

What I hope for

For most of my working life, the work I cared about most didn’t happen during working hours. When I was a researcher at university, the days went to teaching, organisation and industry projects. I’m not against teaching, but it took the time. The research I cared about started at six or seven in the evening, and it was almost never the research that got funded.

That isn’t only my experience. In a randomized experiment with 2,130 evaluator–proposal pairs, Boudreau and colleagues found that “evaluators systematically give lower scores to research proposals that are closer to their own areas of expertise and to those that are highly novel.” Wang, Veugelers and Stephan found that novel papers are less likely to appear in high-impact journals at first but more likely to end up among the top 1% of cited work, after a delay. Katalin Karikó struggled for years to fund her mRNA research and was demoted before it won her a Nobel Prize in 2023. Money goes to what most people already understand, because that’s the safer bet, and that leaves blind spots.

If machines take over more of the average case and make trying things cheaper, there’s more room for work that’s unusual, risky or just something a person wants to do. A machine might cure a disease one day, and understanding how it did that could still fill someone’s whole working life. Nobody would have to do it. You could choose to, join others who care about the same thing, and give a talk about it at ten in the morning because there’s no commute.

So, could AI generate the artist as well? Probably, at some point. That wouldn’t make Graviton Theory less worth making for me, and it wouldn’t decide for anyone else what’s worth listening to. That stays with people.

References

• Arendt, H. (1958). The Human Condition. University of Chicago Press.
• Boudreau, K. J., Guinan, E. C., Lakhani, K. R. & Riedl, C. (2016). Looking across and looking beyond the knowledge frontier: intellectual distance, novelty, and resource allocation in science. Management Science, 62(10), 2765–2783.
• Chess.com (2022). Chess.com reaches 100 million members. 16 December 2022.
• Chess.com (2026). Chess.com reaches 250 million members. 27 February 2026.
• Cowan, N. (2001). The magical number 4 in short-term memory: a reconsideration of mental storage capacity. Behavioral and Brain Sciences, 24(1), 87–114.
• Dinu, M.-C. (2026). Agentic AI for Scientific Discovery. Talk at the 3rd International RL Bootcamp, Salzburg, 18 September 2026. Recording.
• Hayek, F. A. (1945). The use of knowledge in society. American Economic Review, 35(4), 519–530.
• Jahoda, M., Lazarsfeld, P. F. & Zeisel, H. (1933). Die Arbeitslosen von Marienthal. Hirzel, Leipzig.
• Jahoda, M. (1982). Employment and Unemployment: A Social-Psychological Analysis. Cambridge University Press.
• Kakaes, K. (2026). AI has solved one of math’s $1 million Millennium Prize Problems. Quanta Magazine, 8 September 2026.
• Kasy, M. & Lehner, L. (2026). Employing the unemployed of Marienthal: evaluation of a guaranteed job program. Working paper, version of 15 February 2026.
• Keynes, J. M. (1930). Economic possibilities for our grandchildren.
• Maciejewska, J. (2024). Post on X, 29 March 2024.
• Menger, C. (1871). Principles of Economics. Translated by J. Dingwall and B. F. Hoselitz. Ludwig von Mises Institute edition.
• Nietzsche, F. (1883–1885). Thus Spake Zarathustra, “The Thousand and One Goals”. Translated by T. Common.
• Nobel Prize Outreach (2024). The Nobel Prize in Chemistry 2024: press release.
• Nobel Prize Outreach (2026). Katalin Karikó. Women who changed science.
• OpenAI (2026). On the Navier–Stokes Millennium Prize Problem. 8 September 2026.
• Paul, K. I., Scholl, H. et al. (2023). Employment status, psychological needs, and mental health: meta-analytic findings concerning the latent deprivation model. Frontiers in Psychology.
• Sakana AI (2025). The AI Scientist generates its first peer-reviewed scientific publication. 12 March 2025.
• Scientific American (2026). Did OpenAI solve the wrong Navier–Stokes problem? 21 September 2026.
• Shin, M., Kim, J., van Opheusden, B. & Griffiths, T. L. (2023). Superhuman artificial intelligence can improve human decision-making by increasing novelty. PNAS, 120(12), e2214840120.
• Simon, H. A. (1996). The Sciences of the Artificial, 3rd edition. MIT Press.
• Wang, J., Veugelers, R. & Stephan, P. (2017). Bias against novelty in science: a cautionary tale for users of bibliometric indicators. Research Policy, 46(8), 1416–1436.
• Whitehead, A. N. (1911). An Introduction to Mathematics. Williams & Norgate.
• Yoo, C.-M. (2019). Go master Lee says he quits unable to win over AI Go players. Yonhap News Agency, 27 November 2019.