“Grunt work” does not make the researcher

July 27, 2026

“I have suffered, so you should too”

Ok, maybe the discourse is not that blunt (at least not for the majority and not openly), but it is at least

“I have learned suffering this way, so you should too”

The buzzword is “grunt work.” How much grunt work does it take to become a researcher? How many hours of ill-fated coding, debugging, calculations, banging-your-head-against-the-proverbial-wall are appropriate per scientific publication? How much lower will the quality of future researchers be if they do less grunt work than my past self? And how much is our scientific work and training process compromised by outsourcing grunt work to AI tools?

I find these concerns largely overblown. Ultimately, even if we accept the most human-based perspective that our work is primarily about people [1] and not about producing research (a perspective I do not accept, but that is a different story), the emphasis on “grunt work” conflates the goal with one historically essential way to achieve it.

Grunt work is typically defined as some combination of time-consuming, hard, repetitive, mindless, boring tasks. Tracking down a factor of 2 or a minus sign in a calculation or code. Carrying out a long-but-unilluminating, tedious calculation. Pinning down conventions. Discovering new ways to make mistakes that have been made before. The concern is that something happens during the long process of clawing your way out of the pit of confusion that slowly makes you a scientist.

Yes, something happens. But grunt work is not essential for it to happen.

The shrinking role of grunt work

It is a cliché: older generations think the youngsters have it easy. Regardless of whether this is objectively true about normal life in the year 2026, it is certainly true when it comes to technical scientific work. We have access to faster computing, reliable and well-documented code, powerful tools for symbolic and numerical calculations, indexed and searchable information. When was the last time anyone performed a numerical integration by hand? Who was the last person to spend hours tracking down a factor of 2 in their implementation of an FFT, rather than using numpy’s modules?

I find the candid, first-person account of the following discovery illuminating: Saul Teukolsky as a graduate student derived a separable equation that describes perturbations around the rotating Kerr black hole [2]. Its importance cannot be overstated, suffice it to say that the equation has his name. The account of the discovery had two steps. First, a specific technique (the Newman-Penrose formalism, but that is not important for our story) allowed colleagues to make significant progress in the case of the nonrotating Schwarzschild black hole. Schwarzschild and Kerr share the same algebraic structure, so it is conceivable, if not intuitive in hindsight, that the same technique might be successful for Kerr. It was, indeed taking “only a few hours” to achieve a similar result for Kerr, but the equation was still not separable. Second, the equation had to be cast in a separable form, something not guaranteed at all as Kerr has fewer symmetries than Schwarzschild. This took months, including “all sorts of algebraic contortions,” “the mathematical literature on separability of PDEs,” and even a trip to Charles Misner! After six months of on-and-off work “something made [him] try a new set of substitutions” and the rest is history.

This story combines everything: intuition, struggle, expected and unexpected results, and of course black holes! But in 2015, when I was a graduate student in the pre-AI era, this story would have replaced most pen-and-paper calculations with Mathematica. In 2026, in the post-AI era, the six months might be reduced to one day of your favorite LLM trying different “algebraic contortions” that “could have saved [you] a lot of anguish” [2].

Every generation has had it easier than the previous. Every generation has done less grunt work than the previous. If this reduction of friction has made us worse scientists than the previous generations (a position I am inclined to entertain at least on average), this has nothing to do with AI tools.

Support structures

Reducing the friction and grunt work, by whatever tool, will speed up research, increase productivity, improve efficiency. Depending on your perspective, the previous sentence might either make you optimistic about scientific progress or despair at the loss of human skills. Or anything in between. Will the reduced time we spend with our own projects compromise our understanding of our results? And will it lead to less educated, less well-trained researchers? Yes and no. Yes, in the short term of a student project. No, in the long term of a research career. If we keep doing our jobs as advisors or individual researchers.

Long gone are the days when a single scientist or graduate student or Swiss patent examiner worked in isolation, consuming the literature at the rate of a few papers per month and collaborating with colleagues at the rate of one letter at the speed of the post office. Students (and researchers at any level for whom the following also applies) have an army of collaborators, multiple advisors, and the Internet. Now they also have AI agents and LLMs. This entire support structure is already rapidly reducing the time needed for research as we have increasingly outsourced part of the grunt work. The number of authors per paper is increasing and the number of single-author papers is decreasing [3]. But individual publication rates are more resilient [4]: we bundle together and co-author.

This hyperauthorship (my personal record is in the thousands [5]!!) and pressure for students to complete the “training-wheels” projects in their early years before moving on to more complex tasks are bound to chip away at their understanding.

Speaking of, I have a confession to make. My first paper in graduate school contained a section with a short calculation about gravitational-wave polarizations and “null streams” [6]. I was instructed to research the topic by my advisor and, like a good student, I read the literature, did a calculation, made a plot, ran the result by my co-authors, and we included it in the paper. Did I understand null streams and polarizations back then? I thought so. Did I understand them as well as I do now? No, of course not. What would 2026-me think of 2012-me publishing this calculation at my 2012 level of understanding? Nothing good.

When you cannot stop thinking about it

So what changed between 2012 and 2026? I did not keep doing calculations about null streams, nor have I ever tried to debug a code implementing null streams. I have barely worked on polarizations. What happened is that I kept thinking about these topics. In 2015, I read some new papers on null streams. In 2019, I revisited gravitational-wave polarizations from a completely different angle. Throughout, the concepts stayed with me in the background and I would mentally visit them periodically.

Yes, we do need to spend time with a problem to understand it. But that time does not have to be grunt work. It almost certainly should not be busy work. It is time mulling over something, thinking about it, rethinking it, accepting an answer for now but noting a sense of discomfort, periodically revisiting it and seeing if we can resolve this discomfort. My brain is full of half-understood ideas and results, heard from someone or read somewhere, that I might revisit periodically (and if it is any consolation, the rate at which I am acquiring half-baked ideas is increasing with time). Or I leave them behind and they might remain half-understood for years or for whole careers. Thinking ideally also happens during grunt work but it is a distinct process. The need to separate grunt from busy work betrays this distinction. Thinking does not require grunt work.

If AI makes us complacent when it comes to understanding something, this is symptomatic of an upstream, already-broken process.

At my career stage, i.e., someone who can hire students and postdocs, I keep getting asked what I am looking for in a student. My answer is: someone who thinks. Someone who will not stop at obtaining a result but think about how to interrogate it, devise ways to break it, consider implications or ways to expand it. Nothing in this chain is fundamentally broken by the introduction of AI. But I can add one more step: someone who will keep asking AI for clarifying questions or additional checks in order to ensure they understand the output.

A good starting student will follow instructions and carry a project to completion. A good intermediate student will gain confidence, disagree with their advisor, and come up with their own ideas. The final stage is a student who leaves my office convinced that we have it all figured out, but comes back next week (or next month, or next year) and says that ultimately they were not fully happy with the answer, and they kept thinking about it (and maybe used AI), and now they understand it better.

The expanding role of advisors

In the era of pen-and-paper calculations, library visits, or pre-numpy modules, we (the research community that considers results or the advisor who signs off on a paper or a PhD) could be reasonably confident that students (and more senior researchers) have been forced to spend enough time with their work that they understand it. Vigilance has been rising in importance for many years, whether it is about the crisis of reproducibility or the publish-or-perish culture of fixed-term research contracts.

In a thought experiment [7], Alice and Bob take the form of two graduate students with similar external trajectories but very different internal research workflows. Alice is the typical pre-AI graduate student who only uses pre-2019 technology to mitigate any grunt work. Bob outsources everything to AI. The terrifying implication hinges on the assumption that “Bob's weekly updates to his supervisor were indistinguishable from Alice's.”

For years, Bob can pass their work by their advisor, collaborators, the scientific audience of their talks, reviewers, colleagues chatting during conference coffee breaks. That sounds implausible and very similar to the fallacy on which impostor syndrome rests: no one can figure out I am an impostor.

This is not how the research process should work, either between student and advisor or between collaborators. Both the advisor and any collaborators would catch Bob; if not immediately, then eventually. When students or colleagues bring us a result, we do not automatically applaud and give the green light. We ask questions, interrogate it, devise ways to break it, think about implications or ways to expand it. The careful reader will notice that this process is copy-pasted from above: this is how science is done, whether you physically produced the result or whether you are the co-author. Maybe some researchers do not have co-authors. But all students in the formative research years have advisors and most have collaborators, at least in physics and astrophysics.

In fact, the AI student was very identifiable in a story that started similarly to Bob, but proceeded very differently [8]: the advisor caught him on the second question! This is not because advisors act like teachers in class, asking “gotcha” questions with the sole purpose of assessing understanding. In research we ask questions both to assess understanding (if confronted with a starting student) and, most importantly, because this is how we do research in a collaborative or advising setting. And this happens regardless of the academic level of the person producing the result. Sure, some advisors will be complacent and busy, and some students will not become high-quality researchers, and some results will be faulty. What else is new?

The doomsday scenario is that researchers' skills will degrade so much that they will no longer be able to ask questions and interrogate faulty results. Yes, some technical skills will atrophy; I can no longer do long division and my advisor could probably not compute a logarithm. But the doomsday scenario is not about technical skills, but rather about the ability to ask questions. That, I find a low probability statement. The distinction between researchers who think and ask questions and those who do not appears at the earliest training stage. Advisors and mentors should instill it, even starting from undergraduate research. If you get stuck at any level of a question, think about it. If your LLM spits out something you do not understand, think about it and ask more. If you still do not understand it, it is ok, you have the rest of your research career to figure it out. As long as you keep trying, with or without AI.

A moving frontier

The growing anxiety about the role of grunt work reflects a real shift in what actually is grunt work at this point. The dividing line between established results (where the result has been achieved and the grunt work is documented and “mindlessly” reproducible by anyone, such as a factor of 2 in the FFT) and frontier work (where the grunt work is still exploratory, such as a factor of 2 in a first-of-its-kind derivation) is shifting with every new paper or code release. The past grunt work of numerical integration by hand is now the frontier of complex numerical simulations of merging black holes. The past grunt work of discovering the “algebraic contortion” that separates linear Kerr perturbations is now the frontier of nonlinear effects, tackled with Mathematica calculations involving (literally!) tens of thousands of terms. While accelerated by AI, similarly to everything else nowadays, this is not a new phenomenon. Nor is AI fundamentally incompatible with research or training future researchers.

The time dedicated to yesterday’s grunt work is one path to knowledge since it ensures ample time spent on a problem. But it is only a means to an end: that of continued attention to and thinking about a problem. A researcher is not an encyclopedia that can do everything from scratch. They are a process by which ideas and results are being constantly interrogated and revisited. Either in pen-and-paper. Or in a code editor. Or in an AI chat window.

References

1. David W. Hogg, “Why do we do astrophysics?”
2. Saul A. Teukolsky, “The Kerr Metric”
3. S. Wuchty, B. F. Jones, and B. Uzzi, “The Increasing Dominance of Teams in Production of Knowledge”
4. Daniele Fanelli and Vincent Larivière, “Researchers' Individual Publication Rate Has Not Increased in a Century”
5. LIGO Scientific Collaboration, Virgo Collaboration, et al., “Multi-messenger Observations of a Binary Neutron Star Merger”
6. Katerina Chatziioannou, Nicolas Yunes, and Neil Cornish, “Model-Independent Test of General Relativity: An Extended post-Einsteinian Framework with Complete Polarization Content”
7. Minas Karamanis, “The machines are fine. I'm worried about us.”
8. Crista V. Lopes, “The Anatomy of a Learning Stall”

PS. Of course I used AI to edit and format this text. I neither know nor want to learn how to add references in HTML5.