O, brave new academia!

· 9 min read

AI in researchacademic valuesunderstandingcredit and ownership

To publish, where to publish, what to publish, or not to publish, those are the questions.

See also:

MIRANDA: How beauteous mankind is! O, brave new world / That has such people in ’t!

PROSPERO: ‘Tis new to thee.

— W. Shakespeare, The Tempest1

I’ve written multiple times about the effect of AI on research, especially on mathematics. Because of my new hobby, aka spending extra AI compute on math problems that I just happen to find interesting, I’ve had to think about publishing these results. The first one was the claim that the 3-adic zeta value ζ₃(5) is irrational, which I put out two weeks ago as a simple blog post, since I didn’t know what else to do with it, nor could I verify was the proof correct or not2.

I don’t consider these results research really, because I mostly cannot comprehend the proofs, for they are (currently) so out of my scope. At the same time, I feel that these proofs, if correct (which I hold to be very plausible considering the capabilities of current large language models), are still valuable. So, they are not research in my own sense of the word, and they are not worthless either. In the previous post I wrote that it might be the fate of us humans to start digesting the results proposed by AI, and this one is about the results: where to put them, and what to put out with them. My answer, in short: apart from research, labelled for what they are, and never without a route that a human can actually climb.

Just arcAIve them?

As I mentioned in the ζ₃(5) post, somebody has registered the domains arxaiv.com and arxaiv.org, so I’m not the only one thinking about an archive for AI results (or it’s domain squatting, who knows). My collaborators gave me a hint about a new service called Hexagon, and this is not an advertisement for them, but if I understand correctly, Hexagon is exactly an archive that welcomes also (even purely) AI-written papers. Its announcement says explicitly that it “specifically allows submissions entirely generated by LLMs and even where the submitters claim no human understanding”3. The view behind it, as I read it, is that such results should be available to everyone, and may be correct and of interest, even while they are still outside the scope of human understanding.

One of Hexagon’s founders, Ben Antieau, had himself found a result with AI, on the Hodge numbers of fourfolds, that he refused to put his name to as a paper and, like me with ζ₃(5), first put out as a blog post. Unlike mine, his result was checked and understood by the people around it, and it now sits on Hexagon labelled as understood in full4. Apparently the whole point is that the site wants to be the first point of contact of a new proof with the world, even before it is understood by humans. To me it seems to be a kind of a laissez-faire take on the future of math: let proofs come from wherever, AI included, get them out, and sort out the understanding later based on some criteria to be determined.

However, I’m still hesitant to post them anywhere other than my own site as notes. If I posted them on Hexagon or on any other archive that takes AI papers (don’t get me wrong, that’s not necessarily a pejorative), I feel I’d take on a higher responsibility for the result, even though that’s precisely what I kinda wouldn’t want to do: I want to share if there is interest, not posit the work on the same standing as “actual” research. It would be nice to get some traction for these papers, should they be of interest to the community. In any case, I’ve been thinking about starting to post my vibed works on Hexagon, ζ₃(5) as well as other projects, for example one related to the Hodge conjecture, soon™, mostly with the label that there is currently no human understanding for these results.

It’s complicated

I’ve also been thinking about my research page on this site. I set high standards for what I call research that I have done, and the most important one is that I need to understand every single step taken on the way to a result. Obviously mathematical notes which I give with a “maybe” do not fulfill that requirement. Therefore, the research page now has a sub-page of its own, AI-driven results. The main research page still shows the primary research articles, the ones I truly understand, that is, the ones I consider mine. The sub-page carries a disclaimer that the work on it was done with AI and is not refereed, or even necessarily understood by a human, with a link to this post.

The situation is obviously complicated, with extreme Luddites saying that none of these works should be posted anywhere, and extreme accelerationists saying that everything should be posted freely, so that more AI systems can use the results and get even further results. My position is somewhere in between: publish what may well be correct, but never without the means to check it. It would be weird to keep pushing my agent army to make these results and then write a blog post every time describing how I don’t understand the result and how it feels weird to pose it as research, even though that’s kind of what it is, but then again, it kind of isn’t. So from now on, they go to the sub-page and most likely to Hexagon too, without a post each.

It would be even weirder to keep all of the possibly new and interesting results hidden, simply because the status quo has not yet shifted into a direction I believe it will, giving more way to AI-driven proofs. We are on the brink of a brave new academia and later also a new world, and I feel we need to re-evaluate what we consider research; the old paradigms can’t help us that much further, and this kind of a sub-page for research is a small part of finding new methods to arrange our work5. The “proofs” themselves in my case are notes, I guess, that may never become (pre)prints.

Some assembly required

I’d say the minimal fair way to publish is to at least give the humans a chance, so to speak. (And I don’t want this to sound like some robot saying that “maybe we should also give the humans a chance” :D) What I really mean is that if we as humans publish AI-driven research results, we should give ourselves a realistic chance of understanding the proofs with enough work. Providing such a chance then also enables one to critique these results.

The first part of that chance is a companion note. Its point is quite straightforward: someone with roughly an MSc or PhD in math should be able to understand the paper the LLMs produced by working through the companion and effectively building their way into the proof from the ground up, in at most a few hundred pages. The Leiden Declaration on AI and mathematics of this June, endorsed by the International Mathematical Union, points in the same direction, suggesting that standards for such work “might include requiring human descriptions of central arguments obtained by automated tools”6. Then again, the companion is written by the same machines, so it is a claim too, not a guarantee.

Of course, if you’re a cynic, you’ll say: “C’mon, these companions are made selfishly just for yourself, because you need them.” That is exactly correct: I’m compiling these companion notes also for myself, to understand what the proof is about, but I’m not sure they will be enough for me; I might need something a bit more explicit and extensive, if I even have time to go through any of this while keeping this a hobby. However, publishing them with the paper feels like the least I can do to make the proof more accessible.

The second part is the code: the minimal set of programs needed to check the result, not everything the agents produced along the way, just what it takes to reproduce the checks. I did not originally do this for ζ₃(5), but its code is now public7, and the note says where to find it. Not just so that somebody can reproduce my results by giving similar prompts to an AI, but so that anyone can rerun the checks immediately. That is the true open-science pathway, the true ideological belief that openness is what brings this society forward. (We should at some point also discuss open science and the openness of society as a whole, now that AI opens up so much that used to be closed off by effort or expertise, even though the models doing the opening are themselves closed.)

In the previous post, Laissez-faire, I wrote about Tao’s lecture “Math 2.0”8, and about if the human-centric view of math could be somehow kept even in the age of AI-driven mathematics research. Considering what Tao discussed, what I’m doing here, publishing results without understanding them, seems to be precisely what he doesn’t want done with math and AI. I can’t give the associated seminar talk on these papers, so what I put out are “incomplete proofs”, labelled as such, together with the means to complete them, in the hope that someone does.

So the research page now points to the AI-driven results, which opens with a short statement and a link to this post, and that is where my purely AI-driven results go from now on. These results still won’t be considered primary research, as I said. They are AI-driven results that I have tried to make understandable to anyone willing to put the work in. We need to recalibrate ourselves, and testing things like this sub-page and its rule of note, companion and code is how we find our position as scientists again: by experimenting with what feels right in this new state of things. I hope this post brought some clarity, and if you’re interested in these results, please go ahead and formalize them in Lean, take the code, work through the notes, do something fun with them!


  1. W. Shakespeare, The Tempest, Folger Shakespeare Library edition, ed. B. Mowat and P. Werstine, https://www.folger.edu/explore/shakespeares-works/the-tempest/read/5/1/. ↩︎

  2. ζ₃(5) is irrational…maybe. ↩︎

  3. B. Antieau, “Hexagon”, guest post on T. Tao’s blog What’s new (6 October 2026), https://terrytao.wordpress.com/2026/10/06/hexagon/. ↩︎

  4. Hexagon, hexagon:2609.00104, https://hexagonmath.org/abs/2609.00104v1. ↩︎

  5. Ever thought you wanted to feel what a Kuhnian revolution driven by a paradigm shift felt like? If yes, why are you sad when something finally happens? ↩︎

  6. J. Alper et al., “Leiden Declaration on Artificial Intelligence and Mathematics” (2 June 2026), doi:10.5281/zenodo.20302944, https://leidendeclaration.ai/. ↩︎

  7. https://github.com/ahippelainen/zeta-3-5-irrational-public ↩︎

  8. T. Tao, “Math 2.0”, public lecture, AIM Public Lecture Series, Caltech (9 October 2026), slides, https://teorth.github.io/tao-web/slides/math-2-0-caltech-2026.pdf. ↩︎

Comments

Comments run on GitHub Discussions, so you need a GitHub account to post.