← Back to Blog

It Took AI to Reveal Moral Weaknesses in Academic Publishing Behavior

Article Preface:

Thanks for taking a look at this writing. As I hope I make somewhat clear below, I am not sure if I take my own argument below fully seriously yet. I think writing this piece has helped me think about it more, certainly. In order to really know if this argument is decent, I will need more discourse with others. Also, tell me what you think regarding the parting thoughts.

AI Has Exposed Academia’s Fissures in Its Publication Model

I am finding myself occupying a strange and incriminating perspective these days. And, as usual, what might it be about? That’s right! Everyone’s requisite part-time job: dealing with AI-related news stories. Well, it is not everyone’s new part-time job. It most certainly is, though, an academic’s new job. Yes, that’s right. Academics have all found themselves overbooked now. The aberrant growth of AI participation in their field is like the reticulate bodies of chlamydia in a host cell.

Recently, OpenAI assaulted GitHub with a flood of 722 manuscripts. These documents claim proofs of theorems in a variety of subfields of mathematics. In the press, this day has been deemed the “Mathocalypse.” Of course, we are no longer strangers to the echoic scream of mathematicians “losing hope” in their practice (and astrophysicists). Of course, math as vitality—i.e. actually doing math—cannot be extinguished: Chess players still play chess; Go players still play Go; players still play . And even biochemists still work alongside AlphaFold. So, what’s going on with mathematics?

Practically speaking, of course, we can gripe at the issue of what Terence Tao called “(proof) indigestion.” Once AI generates a proof of a theorem, we require the process of peer review to occur before this theorem is subsumed into general mathematical knowledge. Currently, peer review is still conducted by humans. Thus, to contribute such a high-volume output to a preprint server is to overload the field of working mathematicians with the burden of all this cross-checking. In an extreme and limited example, one can analogize this behavior to receiving a manuscript for a 10,000-page book from an eccentric author who claims that he’s offered a proof for hyper-fractional dimensions of mega-spacetime. Thank you for your… interesting… contribution, good sir. Unfortunately, time eludes me in this reading.

Now, though it might sound strange, I am not so interested in the practical aspect here. In some regard, it’s actually quite boring: We in science are already acutely attuned to the practice of peer review. After all, that is the major barrier to getting our work published in academic journals. While I like Terence Tao’s coinage of the terminology, I suspect that it is not at all even close to what really galvanized his appearances on BigThink and other news outlets. Academics have already “calculated” (anticipated) that proof indigestion is a plausible outcome should tech companies unleash their AI-derived scientific results.

What we must really inquire into is why OpenAI decided to deliver this information in the fashion that they did. Was this a spontaneous decision made in the spur-of-the-moment blaze of passion? Perhaps they stumbled onto all 700 results and were so elated that they simply had to show mathematicians! And that is a reasonable reaction from a scientist who found some incredible result in their laboratory: they’d want to disseminate their result.

That is not what is happening here.

The decision of OpenAI to release this information in the academic ecosystem in the manner they did was easily calculated. We all know why.

The primary route by which academics present their work these days is through publications. Here is a brief crash course for those uninitiated in this practice:

In order to receive not only attention for your work but also future funding, you must routinely procure a list of your work. This work is represented as a whitepaper that finds itself either on a preprint server or in an academic journal. You get major points if it’s published in an academic journal, because that implies several other people scrutinized and vetted your work before the bright green, good-to-go checkmark. However, preprint articles (ones that aren’t peer-reviewed) are frequently listed on academics’ websites or curriculum vitae, and effectively bolster one’s publication record. In fact, a dramatic amount of influential papers remain to this day on arXiv rather than being embedded in an academic journal.

While publications have been the norm for the dissemination of scientific findings, the modern practice has mutated. Your production of academic papers—where papers these days are somehow isomorphic to “deliverables,” a term used often in the corporate world—“proves” your significance to the field. There are specific, so-called bibliometric indices that quantify the “significance” of these papers, the most infamous being the h-index. I do not have enough wind to spare to elucidate these indices, and so we will not continue to do so.

Since the introduction of these metrics induced an evaluative norm of one’s “scientific worth,” several types of “coping mechanisms” (or “strategies” to “game the system,” what have you) have emerged. A legitimate example is to feature on a “hyperauthorship” paper, such as a paper reporting the findings of a large, collaborative experiment. An illicit way is to produce “salami papers.” Either way you choose, the objective is to appear better under the metric than you otherwise would.

For OpenAI to deliver this volume of “scholarship” within the academic ecosystem is actually the result of our negligence in decrying this poisoned practice of publication. It is our fault for having allowed this metric to dominate our evaluation of others. It was our permissive model of demonstrating “productivity” that allowed chlamydial AI companies to bind to our academic reticulum. OpenAI’s behavior was rational choice, in the sense of the economic theory. It was not an aberration. It was not irrational. It was calculated, deduced through logical thought. And we played a role in it.

OpenAI exploited our norms of publication for precisely this objective: to directly compete with us as an effective practicing academic. Its remarkably productive research career has blown all of the rest of us out of the water, the company finding itself soaring to skyscraper h-index levels. (Obviously, it hasn’t done this yet. But having a large Npaper is the first step for a high h-index. See An index to quantify an individual’s scientific research output.) (Human) academics are now compelled to “collaborate” with AI tools for the promise of “accelerating” their research, which practically manifests as a bolster to their bibliometrics and presentation.

What is further upsetting is that there have already been several retractions of the announced results. If a single working scientist had as many retractions in such a short span of time as did OpenAI, it would likely forever hamper their continued scientific practice, affecting everything from personal reputation to hiring decisions. However, for some reason, OpenAI has found a nice wedge to camp out in: Why, currently, do we not conceive of OpenAI as a single working scientist that finds itself as a single (of many) author on a scholarly work? Under the extensional classification of research output—the primary objective of a bibliometric index—the “scientist” OpenAI is just a hyperproductive researcher with a relatively small number of mistakes; in absolute terms, it’s doing just fine compared to the rest of us puny, working scientists!

We academics did not anticipate that this conduct exhibited by AI companies would become a problem. Otherwise, we would have certainly acted to prevent it, or at least regulate it preemptively. Do we think mathematicians now “losing hope” and lamenting the extinction of their careers and livelihood wished this so? How about astrophysicists publishing opinion pieces on arXiv titled Why do we do astrophysics?? Certainly not. Perhaps we took the naïve, now-dated perspective: “An AI that can help me check my work? Sure! I’d love a tool like that.” We are hearing the exact opposite. So, we must ask: When, or how, did our perspective change?

So, in this damning perspective of mine, perhaps academia deserves this misfortune and hijack. If academics really believe in the dignity of human scientists, then they would not only drop these evaluative games they play, they would condemn them. Now, the rejection of shallow interpersonal characterizations does not entail forfeiting the essential elements of meritocracy in science; it is merely a matter of respecting people, their knowledge, and their abilities. It also does not imply the abdication of all research pursuits; we can still do research without the encompassing, enveloping apparatus of this publication model. After all, why do mathematicians do math? Ask them:

“Omdat wij dit graag doen.” (because we like it)

— Stefaan Vaes

“…one could earn one’s living by playing.”

— Pierre Deligne

To denounce a mutated structural anomaly does not (always) necessitate the rapid disintegration of the activity it so happens to institutionalize. Perhaps it is now time to condemn scientists that continue to operate based on these principles; not only are they amoral, but—a reason that defenders of these metrics most certainly can appreciate—they are unscientific!

One major gripe that people will no doubt have is that academics only play these games because they have to: From political to sociological forces, especially nowadays, playing the game is sometimes the best option for keeping your head low. (In some way, this decision-making is also rational choice in action.) Obviously, it is difficult to stand up to people who are otherwise willing to fund your multi-million dollar dark matter search. Normally, I wouldn’t be so agitated by this reasoning. But this situation is different. Why? Because these are academics. They occupy a particularly distinguished class in our society, a class known for being intellectual, i.e. “good at thinking,” or something like that. If we are so “good at thinking,” why did we not foresee this crisis? Why did we not dispense our insulting models of “scientific evaluation?”

Presently, academic papers are still branded with the fiery names of people. Contributions of AI technology are typically featured at the end of the paper. What effect does this have from the extensional, bibliometric perspective? It makes it so that these publications go under the names of people and not of personified LLMs. Otherwise, Claude or Gemini would find themselves an accomplished scientist! (Or so it would seem, right? Remember: the main de-correlation/de-programming we wish to suggest is that publication practices, especially in this day and age, have virtually no grounds for evaluating scientific research.)

The problem that we currently have is the seeming contiguous adhesion of two primary processes of academic practice. The first is the standard work-and-publish sector: Research is conducted, it is written about, it is peer-reviewed, it is published. The second is the discovery of foundational results. Both occur, and that is how science proceeds. The characterization is certainly a distasteful one because it is not particularly nice to make the categorical separation between “background work” and “foundational results.” However, it cannot be denied that indeed foundational results in physics exist! In addition, it is unclear that foundational results were as valuable as they were until quite some time later.

Scholarship and appearance of scholarship are not tightly correlated. Examples abound:

Theoretical physicist Ahmed Almheiri has 26 total papers in his entire career, spanning (as far as I can tell) 15 years. But look at how frequently his work was referenced: 15 citations, 1,811 citations, 574 citations, 32 citations, 781 citations, 899 citations, 130 citations, 63 citations, …, 1,133 citations, 923 citations, 426, 301, 1,041, 69, 924, 36, 27, …, and so on. There are some years when he published nothing. Edward Witten is of course the classical example, the “Mozart of Physics,” who not only dominated bibliometric indices but whose contribution to physics knowledge is indispensable: 344 citations, 217 citations, 701 citations, 746, 211, 179, 113, 274, 1,072, 95, 355, 991, …, and so on. Grigori Perelman contributed a proof of Thurston’s Geometrization Conjecture with two arXiv submissions—they still sit on arXiv to this day (i.e. they are “unpublished”). His work provided foundational understanding in incredible fancy math stuff that I do not myself fully understand (except that it’s related to differential geometry). This scholarship that endowed him for a Fields Metal—which he rejected. And, in principle, every time we write down Maxwell’s Equations, we ought to cite James Maxwell, probably along with André-Marie Ampère, Oliver Heaviside, Michael Faraday, Carl Friedrich Gauss, Charles-Augustin de Coulomb, along with many others. I think perhaps some of these giants would have disgraceful bibliometrics.

Now, as I commented on earlier, foundational results do not pop out of a vacuum like a quantum fluctuation; it often takes the “background” work that is always whirring before we stumble upon these awesome ideas. That is why I mentioned the two processes. Similarly, I do not wish to fully strike against the demand that our work manifest as published documents—and frequently so! After all, written documents provide a compact form of our findings that are easily exchanged with other researcher. In addition, as I will comment on momentarily, I do believe in the notion of deadlines as encouraging accountability! However, what we must attend to is the likelihood the metrics we use to evaluate working scientists do tremendously more harm than good. Soon, perhaps, AI companies will appear as some of the most productive and enthusiastic researchers in mathematics, physics, and astronomy (among others). What looks like “accelerated discovery” is in fact just accelerated production of the first type of process in science research: the hum, the buzz, the background work that often goes unnoticed. Do we really wish for this possibility to occlude our horizons? Remember: they will only appear as such. But the impact they have on our fields, in a sociological scene, will be real.

Let us conclude with a happy suggestion. I propose that academics ought to be afforded a two-month “holiday” in which they are free to conduct and think about any research they like. This holiday may be like a “gift” to those intellectually ambitious enough to step foot into the vicous pursuit that is academic practice. I also believe, because of a very separate conversation that we did not touch on here, that this practice may restore some equality to academia.

Why only two months and not longer or shorter? I don’t think it can be longer for a few reasons. Firstly, I actually am somebody who believes that deadlines are actually good (for the most part). Secondly, I also believe that receiving funding for scientific research ought to come back around in some facet. But it cannot be shorter because thinking about enthralling ideas and research directions requires time to elapse. It requires having the idea, letting it marinate in mind, trying something out and failing, taking a walk around a lake, and then returning and making progress. It requires, as Edward Witten once said to Brian Greene:

“But I can assure you that there’s a lot of the time trying to do research in theoretical physics feels like you’re hanging around and doing nothing.”

And that’s Edward Witten saying that—whose h-index is higher than yours.

I think that, in fact, some of the most valuable steps forward in science may arise from this two-month playful interlude. All that’s missing now is a name for these two months… Anyone have any ideas?

Notes/References:

  1. “Everyone’s requisite part-time job: dealing with AI-related new stories.” I think this is an unfortunate truth. Terence Tao is now beholden to appear on every podcast, every talk show, and every news segment to describe how the field of math is changing instead of actually doing math. As I tell my friends, Terence Tao is (maybe unfortunately) the sort of “CEO of math.” The first fact anyone knows about him is that he was a super genius at math at age 7 while the rest of us were gnawing on Legos. His child prodigy saga likely fueled sustained media presence, and his wondrous achievements in mathematics are a function of his brilliant mathematical insights and intuition. If there is anybody who is upset about being unwillingly recruited to work part-time on AI-science coverage, it is Terry.

  2. “Recently, OpenAI assaulted the renowned preprint server arXiv…” On OpenAI’s math GitHub repository, the company reports, “The current catalogue contains 722 manuscripts organized into 373 families” (as of this writing).

  3. “In the press, this day has been deemed the ‘Mathocalypse.’” Lots of news stories use this word: OpenAI sparks ‘mathpocalypse’ as it publishes full or partial solutions to 372 outstanding problems, AI eats science 1: the mathpocalypse, The Mathocalypse,

  4. "… we can gripe at the issue of what Terence Tao called “(proof) indigestion.” For the original source, consult The paradox at the heart of AI and science | Terence Tao (timestamp included). For previous discussion on this blog about this issue, see We Must Decide if It is Two People or Three.

  5. “In fact, there have already been several retractions of the announced results.” OpenAI retracted three of their manuscripts due to… a sign error (see OpenAI withdraws three preprints a day after releasing 722 manuscripts on unsolved math problems). You know, it is a common joke in physics (probably in math, too) to go up to somebody grinding away at a tough calculation and say, “You missed a minus sign!”

  6. “However, preprint articles (ones that aren’t peer-reviewed)…” arXiv articles are not peer-reviewed but they are moderated. See arXiv.org for more information about arXiv.

  7. “OpenAI’s behavior was rational choice, in the sense of the economic theory.” This alludes to the Rational Choice Model of economics. Most definitely, OpenAI has maximized their utility!

  8. “A legitimate example is to feature on a “hyperauthorship” paper…” It appears that the term “hyperauthorship” comes from Hyperauthorship: A postmodern perversion or evidence of a structural shift in scholarly communication practices? originally. It is described as “massive coauthorship levels,” and, in context:

…we need to consider how multiple authorship, in extremis—what I have chosen to term ‘hyperauthorship’…

I found the term from The h-index is no longer an effective correlate of scientific reputation.

  1. “An illicit way is to produce ‘salami papers.’” A “salami” paper is described in arXiv’s blog post Fair Moderation, Equitable Access, and AI: arXiv’s Updated Rate Limit Policy:

Our moderators are observing an increase in thin papers of narrow scope, as well as ‘salami’ papers, where a single work is broken up and submitted as a set of smaller papers.

  1. “‘Omdat wij dit graag doen.’ (because we like it)” indeed comes from Stefaan Vaes, but I found this quotation and the next one in the article Knowledge Collapse by Michael Harris.

  2. “Presently, academic papers are still branded with the fiery names of people. Contributions of AI technology are typically featured at the end of the paper.” In ten billion new papers, we now almost always see something to the extent of “Awesome AI model [xxMod3lNam3xx] contributed to the checking of the calculations,” among many other things.

  1. “(Human) academics are now compelled to ‘collaborate’ with AI tools…” Look. All I can say is that Andrew Strominger has appeared on an arXiv preprint listing one of his (three) affiliations as OpenAI… See Single-minus gluon tree amplitudes are nonzero. Then, much more recently, his affliation is exactly listed as OpenAI: See The Conformal w -Algebra. That was on September 11th, 2026.

  2. “…this practice restores some equality to the field of academia.” That is because I think that the contemporary (mutated) academic practice we’ve been ranting about here actually beats down on underrepresented folks.

  3. “Theoretical physicist Ahmed Almheiri has 26 total papers in his entire career…” Take a look at InspireHEP, the high-energy physicists’ analog to arXiv: Ahmed Almheiri.

  4. “Grigori Perelman contributed a proof of Thurston’s Geometrization Conjecture with two arXiv submissions—they still sit on arXiv to this day.” The two papers are:

  1. "Edward Witten is of course the classical example, the ‘Mozart of Physics’…’ Terence Tao is often heralded as the “Mozart of Math.” I don’t entirely remember how this came to be. Nevertheless, if you Google "mozart of math", he will be the first result. On his Wikipedia page (Terence Tao) and press Ctrl + F, you will (as of now) find five articles that use these words.

  2. “…as Edward Witten once said to Brian Greene…” See String Theory in 2037 | Brian Greene & Edward Witten, timestamp included.

Updates:

  1. 2026-11-02: I said earlier that “OpenAI assaulted the renowned preprint server arXiv,” but that’s not right. The manuscripts haven’t been released on arXiv yet. They are only presently on their GitHub. In fact, they couldn’t have been legitimately hosted on arXiv according to the article referenced in Note #8 above: Fair Moderation, Equitable Access, and AI: arXiv’s Updated Rate Limit Policy. In that article, it is stated that:

arXiv now limits submitters to up to two submissions per calendar month, with a limit of three total active submissions at any given time.

Coffee Cup