Several people [1], [2] worry that with the advent of Artificial Intelligence (AI),1 research (as we know it) is over. Karu Sankaralingam, a Principal Research Scientist at NVIDIA and Professor at UW Madison, wrote an article with the title: “What Remains of Research When the Machine Can Ideate?” The main thesis is that ideation has been the essence of research. As a researcher of any rank, to ideate is not “another thing you gotta do”; it is the thing you want to do. In some sense, you do everything “you gotta do” so that you can have the time and space to ideate. Folks become professors, and get meager salaries, and live in crappy apartments in depressing cities like Champaign-Urbana, and spend the largest chunk of their time writing the most boring prose one could possibly imagine (grants) in order to win for themselves a little bit of time to ideate! In short, the narrative academia uses to lure people in is: come here to ideate (whether this narrative is realized is a whole other discussion). Ideation is a reward promised to researchers.

But it is also the means to other rewards. More specifically, ideation is—well in romantic people’s minds at least—what creates careers. The first question your advisor, the hiring committee, and the conference reviewer will ask is: what is novel about this work? Novelty, some researchers say, is the currency of the field. But novelty is nothing but the result of ideation “done right.” After all, everybody has ideas. But which of these ideas are new? That’s the first question anyone asks, before they even ask if they are good.

Given the central role of ideation, then, Sankaralingam asks this pressing question: if AI can ideate, then what is left for us, humans? Perhaps more importantly, we are headed toward a direction that will hurt primarily the most vulnerable!

You see, contrary to what many people outside of academia may think, it is not professors who do most of the ideating. They do all kinds of things, but in most of them ideation is almost incidental. This is because their time and incentives leave little space for ideation (e.g., in most cases, to be successful, they need to have multiple projects running in parallel). Who ideates a lot? PhD students. Yannis Smaragdakis, a senior professor at the University of Athens, supports this position. In a talk titled Why do a PhD? given at SPLASH 2019, he said the following addressing PhD students (21:00 onwards):

You are in the ideal position to make significant contributions. It pains me to say it [...] but professors are not in such a position nearly as much as PhD students. A professor is much more distracted in terms of time commitment. People who start their PhD; they have the time, they have the energy—and the courage many times—to go deep [.] People with more experience [...] do not have the same luxury to go deep.

The common denominator to all this is ideation. How do you make a significant contribution? By ideating! Probably you also have to study and understand—probably a lot, actually! But comprehension alone will not get you anywhere. Similarly, going deep alone will not get you anywhere. All that is just a means to effective ideation. Smaragdakis is also talking about having the time, the energy, and the courage. Well, to do what? To ideate! I mean you also have to implement, and read, and understand, etc., but at the end of the day, you need to have the time, the energy and the courage to ideate enough, actively, and daringly. So, as Smaragdakis says, PhD students are in a much better position to do all that and so it is not surprising that it is actually mostly them who do it.

Coming back to my previous point, the fact that AI can ideate is alarming primarily for PhD students. For as a PhD student, that is basically all you’ve got! You don’t have influence/status/power/connections, you don’t have money, you don’t have citations, you don’t have tenure, often you don’t have citizenship in the host country (read: even less bargaining power with your bosses), and you don’t have basic labor rights. I mean yes, every start of the semester the HR will send you this email warning you that oh my, you should work for up to 20 hours per week. But that’s a joke. You know it, they know it, your advisor knows it; and everybody knows that everybody else knows it. Frankly, you don’t even have a standard of living that can be reasonably called dignified. Thus, what on earth is going to keep you going?

If you ask professors—and I have—most will tell you that the PhD is but a precursor to your “actual” life, which I guess is either a tenured faculty position or an industry job with baskets of money. I’m not going to get into the myriad moral problems with this line of thinking. The point is that according to this narrative, the only thing you have had up to know to get you successfully through your PhD and into your “actual” life is ideation. If machines can do that, what happens with the PhD students? What will they do?

—

Levente Littvay, a senior professor from Hungary, contributed to this discussion with a recent article. What makes his perspective especially interesting for this article is that he is a professor in Social Sciences, i.e., on the “other side” of Sankaralingam. Yet, he voices similar concerns. In his article, Littvay describes how the new Claude edition nailed multiple tasks that academics used to do, and did them pretty quickly too. As a consequence, he wonders:

In a world where one can rewrite a grant, write three papers, and a book in 3 days with AI, what is left of an academic’s job?

Littvay’s worries align with those of Sankaralingam, in that he is mainly worried about junior scholars:

What I did with papers 2 and 3 above was to replace the more junior collaborator with the AI. The process was not at all different in anything but speed. Also, I can’t say I enjoyed the collaboration as much, and, yes, what I did took an opportunity away from a junior scholar.

Both Sankaralingam and Littvay have solutions to offer. Sankaralingam essentially starts from a core insight: doing is not that important anymore, where by “doing” we mean implementing, writing, etc. So, now that AI can do a lot, Sankaralingam suggests that “intellectual leadership now” is about skills such as:

• Taste at scale. The ability to evaluate many AI-generated ideas quickly and identify the genuinely promising ones.
• Narrative Construction. Weaving results into stories that matter to the field.

The second point is particularly worrying both for me and Sankaralingam because it implies that “what is left” for researchers is salesmanship. Now, this is in my opinion what everyone except for PhD students is already doing for the most part (e.g., professors of all levels). That’s because PhD students (and undergraduate researchers even more) are the lowest level of the hierarchy: they do whatever needs to be done. Conveniently, they do that by getting paid a shitty salary, or in the case of undergrads, not at all! With the advent of AI, one may predict that PhD students will become salesmen too. However, the problem is that once you don’t need the humans in the lowest level of the hierarchy to perform the tasks they used to perform—because the higher levels can do them at a much lower cost using AI—then you don’t transfer the tasks of higher levels further down. No, you probably just remove the lowest level altogether. As such, the result may not be that PhD students will become salesmen but that PhD students will be no more.

Littvay proposes slightly different solutions. For example, he claims that:

[W]e need a system where less is more. One admittedly ugly way of thinking about scientific contribution is to ask how much we contribute to the training data. If our study adds little new information beyond what the AI already could have derived from its training, we are not really contributing.

Then, he touches upon what is, in my opinion, the most important point: “Let’s start rewarding hard work, not just output.” This brings me to my contribution to the discussion. I have been thinking about it for months, and when I read Littvay’s article, I was hoping that someone finally got to the bottom of it. In my humble opinion, he did not, so here is my attempt.

I think that both Sankaralingam’s and Littvay’s solutions are good, but short-term. You see, if everything goes to plan, AI will eventually become just as intelligent as humans (and then some). That sounds striking but why should we believe that?

First, tools like AI Scientist have literally set this as their goal. AI Scientist's subtitle is “Towards Fully Automated Open-Ended Scientific Discovery” (my emphasis on “Fully Automated”). AI Scientist has already made experiences like Littvay's obsolete. A fellow Computer Science researcher recently told me that he was able to finish a paper that would normally take 2 years in one month! Then, on the more theoretical level, Philosopher of Computer Science William J. Rapaport has argued in a long paper there is no logical reason to believe that AI will not succeed, where success is reaching human-level intelligence. For our discussion, this entails full automation of research. It really does not matter whether AI will achieve human intelligence, or human-like intelligence. What matters is that it will eventually reach human-level intelligence. If you are interested in the distinction, I have analyzed it in a previous article. But for practical purposes, the point is that AI will eventually reach a point in which it will be able to perform most research-related tasks at the level humans do today, and perhaps better. In other words, it will be able to provide equal or more utility to a human.

This wouldn’t be a problem if it weren’t for the nature of the research world: it is fundamentally utilitarian. At the end of the day, what a researcher gets hired for and what she gets paid for is utility. And utility, for universities, is tied in one way or another to money. Why do professors want many publications? Because publications are required to get money (grants), which is what the university wants and expects. PhD students are the workforce in this money-making machine—since they create the publications—and professors are the managers.

Thus, no top university or industry lab works based on abstract principles. No one gets paid to do “the right thing” or “their duty”. Karikó Katalin, the Nobel-prize laureate, was almost fired because she was not bringing enough money for the university. Katalin did serve the core principles of science like: honesty, rigor, integrity, hard work, curiosity, self-discipline, creativity, novelty, courage, determination. She did everything which the research world proclaims to consider “right” or “the essence of research.” Yet, she was eventually demoted (after trying hard not to get fired) because she did not have any grants to show for it. In a similar manner, what do advisors expect from their PhD students? In fact, what are the requirements to get a PhD? A certain number of papers. It is not principles. You get paid to bring utility, period.

In addition, the roots of the modus operandi of the research world are deep. Consider Science, the Endless Frontier, the masterpiece essay by Bush that is the primary reason behind the structure and operation of scientific institutions in the U.S. today. This is a quintessential example of a utilitarian text (albeit at times in disguise). Or consider liberalism, one of the central tenets of many scientific institutions today (although peer review is clearly non-liberal). These institutions understand liberalism the way J. S. Mill envisioned it: we should support the marketplace of ideas—i.e., the unencumbered exchange and dissemination of ideas—not because it serves freedom or justice, but because it benefits the utilitarian calculus.

This is why Littvay’s “let’s start rewarding hard work, not just output” will never work as long as the research world operates the way it does today, and I don’t think it will change any time soon because academia is, operationally and bureaucratically, one of the slowest-moving institutions one could ever find.

In a similar vein, we can understand why all the other solutions presented above are temporary. For example, Sankaralingam says that one should have “[t]he ability to evaluate many AI-generated ideas quickly” or be able to “[weave] results into stories that matter to the field.” AI will be able to do these soon, if not already. Similarly, Littvay writes that “[i]f our study adds little new information beyond what the AI already could have derived from its training, we are not really contributing.” But AI can mostly already arrive at results beyond its training, and it is about to get a lot better.

Thus, if we are trying to come up with ways of how humans can still provide utility in research, then it will be gameover very soon. Of course there are things here and there that bring some utility and only a human can do—like changing a graphics card or supplementing a conference talk with a beautiful physical presence—but the human will not be the protagonist. “We will be better on a sufficient number of things that (currently) matter” is not on the menu, sorry.

In a short while, a whole range of activities—coding, writing, thinking—when done by humans, will simply not be able to outperform AI in utility. Thus, these processes as human activities will disappear. That is, coding itself will probably not disappear. But coding done by humans—or at least with humans as the protagonists—will disappear.

It is sometimes hard to confront that reality because our environments are arguably misleading. Even though employers do not care at all about this or that human activity, at times they act as if they do. For example, when you get hired at a big company, they give you a great desk and a huge-ass monitor and they organize “team-bonding activities” and they offer you lunch. Understandably, these “gifts” may make you think that they care about elevating your activity. But no, the reason you get all these goodies is because they think you will produce more utility that way.

The same is true for advisors, PhD students, and papers. At the end of the day, your advisor does not care whether you or Claude does the coding. She does not care if you work from home and you’re totally alone or if you come to the university and build relationships. What she cares about at the end of the day is publications. The reason she may urge you to come to the university to “build relationships” is because in her mind, relationships lead to publications. But as we said, AI will be able to create publications almost on their own pretty soon. Thus, if we keep treating PhD students as paper-producing machines, we will see the end of research by humans, because real machines will become better.

Sankaralingam describes the big transformation that happened in chess in the last couple of decades:

Twenty years ago, if you told chess grandmasters that a laptop would soon beat the world champion—not occasionally, but trivially, every single time—they would have found it difficult to believe. Chess was the quintessential intellectual game. It required creativity, intuition, the ability to see patterns invisible to ordinary minds. It was, in some sense, a proof of human cognitive exceptionalism. And then it wasn’t. Today, Magnus Carlsen—arguably the greatest chess player in human history—would lose to a chess engine running on your phone. Not sometimes. Every time. The gap isn’t close; it’s embarrassing. A $50 piece of software plays chess better than any human who has ever lived or ever will live.

In a utilitarian context, that would be the end of chess as a human activity. But it wasn’t, as Sankaralingam notes:

What happened to chess culture? It adapted. Humans still play each other, and we find it meaningful. But no one pretends anymore that human chess represents the pinnacle of chess capability. The locus of “best chess thinking” moved from human brains to silicon, and it’s never coming back. Are we watching the same thing happen to research?

This is factually true, but I think there is something more to it. Indeed, all contemporary debates at the time when these events were unfolding were about what Sankaralingam identified: chess capability. But in my view, capability was never the point of chess; it was as tangential as in any other game. And the reason it was and is tangential is because chess was never about utility. In fact, I think we can all agree that a chess game has no utility at all! Technocrats would call it useless. Thus, at the end of the day, it simply does not matter whether a machine can do it better. It is philosophically and historically interesting, but it will not put an end to the human activity of playing chess. The activity changed of course; for example, humans started learning from machines. But I do not think there was ever any serious discussion—among philosophers, scientists, and chess players alike—about whether humans will keep playing chess.

To even question that would seem as preposterous as declaring that you will stop playing football because you will never be as good as the top players, or even good enough to get paid for it. That is, some other agent scores much higher on capability than you will ever do. So? Capability is not the reason you play football. Similarly, of course it feels good to know that you are good at sex, but this is not the main reason you have sex! Humans play chess and football and have sex not because of their utility, but because they enjoy doing these things. More schematically put, none of these activities is treated as a means (to utility); they are ends themselves.

Thus, to come back to Sankaralingam’s question: “Are we watching the same thing happen to research?” No, because research is not like chess. Research as an institution, in the last couple of decades, has always been about utility. That was not true for medieval scientists or Aristotle, or even the mathematicians who laid the early theoretical foundations for cryptography (but who at the time had no idea that their theorems could ever be used for anything at all). But it is clearly true today. Thus, when a machine becomes better, the institution as we know it will throw the humans out of the window. They will not go away immediately, because as we said the research institutions move at snail speed, but it will eventually and unavoidably happen.

The only long-term solution is to completely rethink research as a human activity. If we believe that there is value in research being done by humans other than its utility, then we need institutions that operate based on that. Littvay’s “let’s start rewarding hard work, not just output” is but one manifestation of this core principle.

In fact, AI has the potential to create a positive effect (although I do not believe it will happen). It has the opportunity to make research a hobby—just like chess—left for only those who really care about the activity and its humanitarian value. This of course means that in the current (utilitarian) economic system, the researchers will not be making money out of this activity. But this is precisely why only those who like it will do it. As a bonus, I think that many of the power structures will get destroyed exactly because most of them rest on utility.

In the long run, we may see the true “medieval” research re-emerge, as now careers will not be tied to papers or any other utilitarian measures. We may even finally see the end of peer review as we know it, as now people will simply not care what Reviewer #2 thought after reading the paper for 1 and a half minutes. They will send the paper for review to people who can actually engage with the work in a carefully critical manner, or they will simply publish it online. Because no one I know would choose peer review as we do it today if they didn’t have to. Stay tuned, this is the topic of an article that will appear soon.

1. The capitalization matters. It signifies that AI it is simply a proper name given to this type of software. Using lowercase would imply the claim that what I refer to as “Artificial Intelligence” is a form of intelligence. See What is Artificial Intelligence? for a discussion.↩