Our guest column is a call to action in this age of LLMs, from Po-Ling Loh, Professor at the University of Cambridge Statistical Laboratory. She explores the urgent threat that AI poses to mathematical statistics… and proposes a solution that could lead us from competition to collaboration.
AI has driven a dagger through our scientific profession. This is now irrefutable: Three months ago, debates were still taking place about whether ChatGPT was good enough to solve research-level math problems. Maybe it was just combinatorics or number theory on the chopping block, while mathematical statistics would remain relatively safe due to an enlightened combination of human-inspired, science-driven insight. Could AI actually be utilized effectively to check proofs, while the process of mathematical formalization lags behind? We now see examples announced every week showing that state-of-the-art LLMs can indeed be harnessed to make substantial progress on theoretical problems in all areas of mathematics. And if you are still a skeptic, I encourage you to find a friend with a GPT Pro account and ask it a question you are pretty sure AI can’t answer, and be astounded by the response.
I chanced upon the power of AI tools for math research in mid-April, soon after GPT-5.5 was released. I had previously spent at most two hours using GPT, despite starting a $20/month subscription last summer: When I used GPT-5.3 to assist me with proving a lemma in February, I found it helpful in rapidly reducing my search space and impressive in its knowledge of what I felt were esoteric topics in my field. However, GPT made silly mistakes, declaring that a convex function was concave, and proceeding to “prove” the opposite inequality that I wanted. Nonetheless, after several iterations, GPT helped me craft a proof that might have otherwise taken a week. When I realized that GPT-5.5 was so much better, compounded with the public announcement of GPT’s solution to an Erdős problem in mid-May[1], I was initially alarmed. However, my alarm quickly transitioned to reassurance: Although my path to becoming a professor was motivated by the thrill of doing math research, if research were phased out of my profession, I would still derive a fair amount of satisfaction from teaching—and there will always be a need to educate the next generation of students, even if the form of that education may be overhauled by LLMs.
Nonetheless, my complacency soon yielded to deep unease. Something I take as seriously as my personal engagement in research is mentoring students and postdocs. Junior researchers who have worked with me can attest that I go to great lengths to explore all the paths a mentee might take and carefully discuss the options with them, before accompanying them down the path they ultimately choose. What goals should a student or postdoc be setting in this new era? In May, GPT-5.5 made it incredibly easy to generate new, complex, rigorous, mathematically sophisticated papers in a matter of days. By July, GPT-5.6 was accelerating the process to a matter of hours or even minutes. In the academic world, where evaluation metrics are defined by publication speed and volume (subject to appropriate quality standards), would we soon be seeing the bar for paper acceptance rise infinitely high? And would the need to compete on such massively accelerated timescales necessitate reliance on premium AI tools, creating a sense of despair and futility as we attempt to compete with machines? The fact that I could not envision a future path for my mentees in academic research gnawed at me in a way that even an unsolved math problem could not.
I began to talk to more people—at Cambridge, outside Cambridge, in statistics, in pure math, and applied math. I tried to ascertain what plans were taking shape within the IMS, first regarding AI policies in publications (since the conversation in the Cambridge math community, perhaps heavily influenced by the Leiden Declaration[2], centered around the fact that we need to save the ethics of our field)—and then more broadly, about long-term plans that could help us survive in light of AI. I met sympathy, concern, and despair, but a disappointing lack of zeal for action. My original suggestion was to have an open, distinctly multigenerational discussion about the effect of AI on our field in the statistics community[3], in which young researchers could play a critical role in shaping whatever new policies we might develop. Even if it takes time to develop sound policies, I think we do our junior researchers a disservice by pretending that the status quo is possible and everything is under control. Instead, should we not create a space for the exceptionally creative individuals in our community to exchange ideas about how to save our field?
In late July, I co-organized a small statistics workshop in the Black Forest, where I led a discussion on AI. How much has AI been affecting people’s workflows, and where do we think we are headed? I was surprised by two things: nearly everyone used AI frequently, and, despite this, most people classified themselves as optimists. However, during the discussion, it became evident that nobody, even the optimists, had a plan. Many people assumed it was just a matter of time before the journals or statistics societies adopted new guidelines about ethics and evaluation. When I told them that there were scant plans in the works, the response was that “someone higher up” should take action. I pointed out that our group of discussants included influential statisticians—if not us, then who?
“Hope is not a plan, and nostalgia is not mentorship”—Ken Ono
As I broadened my search of what other mathematical communities are doing about AI (the SIAM community had a discussion about the impact of AI at their main conference in July[4]), I came across the opening lines of an address by Ken Ono, a prominent number theorist and mathematical influencer, to the National Academy of Sciences[5]. He was recently appointed to head a committee to plan the future of math in light of advances in AI. In the address, he talks about how our PhD students entered one profession and are graduating in another, and we have a responsibility to the next generation to help plan and shape the field. This greatly resonated with me, and I reached out to Ken (I knew him from when I was in high school and he was an assistant professor at UW-Madison, but I had not spoken to him in at least 12 years).
I was hoping the conversation with Ken would generate some actionable solutions. However, Ken told me that in his opinion, the best way to train current math PhD students would be to encourage them to explore other fields and expose them to non-academic careers. He predicted that in the next three to five years, theoretical math research would be reduced to a small fraction, and the remaining academics would function more like humanities researchers, expositing the beauty and truth of mathematical arguments rather than working actively on proof generation. He assured me, however, that there are other areas of the applied sciences which cannot be “solved” by AI, hence there are other academic fields one could certainly pivot toward where real human-led research would continue to happen.
I agree that other areas of science exist for which AI serves as an increasingly powerful, incredibly helpful tool rather than an existential threat. Perhaps changing the direction of one’s profession toward these areas is the easiest, cleanest solution. However, my own previous attempts to engage in more applied areas of statistics have always led to me gravitating back toward theory. If I were to pivot away from my current profession in mathematics, there are several non-academic professions I would more eagerly embark on than remaking myself as an applied scientist. Many of my current and former group members, and others in my direct collaboration network, are similarly passionate about their research due to the abstract, problem-solving aspects. So where does one go from here?
In light of these countless conversations, and in the absence of having found any solutions, I would like to propose a radical one of my own. It is a proposal that gives me a glimmer of hope for a future where the problem-solvers among us have not been forced to jump from a rapidly sinking ship, but have been steered into a safer harbor.
A new publication model
What if we were to create a new publication model that proactively incentivizes human collaboration? While LLMs have a dizzying ability to solve math problems and improve upon existing work, many mathematicians left alone with an LLM—without the external pressure of publication or competition—actually find the process of making rapid advances on a research problem exciting and exhilarating. The issue is when one must package these investigations into a paper, knowing that someone else could “one-shot” and strictly improve the paper within a few hours, with the help of an LLM. However, an integral part of one’s experience as a mathematician comes in communicating, discussing, and unpacking the ideas underlying a proof[6]: An interesting idea that one researcher may have generated (possibly with the help of an LLM) would be received enthusiastically by other researchers in the same sub-community, and subsequent discourse would serve to refine the result. Rather than accepting a new reality where researchers are disincentivized to share their work publicly through an arXiv posting or research talk prior to publication, we could purposely excise speed and competition from the research pipeline, thereby stemming the cutthroat competition that threatens to engulf our field. In such an alternative reality, I believe mathematical joy, depth, and beauty could continue to flourish.
We should not overlook the fact that the sharpening and improvement of research ideas by other human beings is entirely in peril[7]. Conventional peer review provides an avenue for other researchers to offer feedback on a paper. However, LLM usage is becoming increasingly prevalent in the revision process—a particularly alarming use of LLM vs. LLM has played out in the recent author rebuttal period for NeurIPS 2026. AI models tend to over-indulge a user’s ego, and if we continue with the current publication models, it is becoming increasingly unlikely that an author would ever receive actual human feedback on a paper as part of the review process. Looking ahead, even if the mathematical abilities of an LLM may soon surpass those of an individual mathematician, we can hold onto hope that the collective skills of a community of mathematicians, collaborating together and assisted by LLMs, might achieve mathematical advances beyond what one human-
directed AI system could discover.
There are elements of my idea already present in the community. From time to time, groups of researchers convene for topical workshops at institutes such as Oberwolfach, Dagstuhl, Bellairs, or the Simons Institute and collectively write papers that are a result of a productive week spent together. What if this became the norm? The Polymath Project, an experiment which was initiated in the pure math community almost 20 years ago, led to several breakthroughs in mathematics[8]—perhaps the arrival of LLMs will finally provide the impetus to revisit the idea in earnest. This model would eliminate the competition that has crept insidiously into our field over the past few months and indeed enhance the importance of proper mathematical communication and the development of deeper understanding, which everyone agrees is the human aspect of mathematics that LLMs are less likely to replace.
Practically speaking, the new publication model I am proposing would only consider papers that are a collaboration of enough experts in a sub-area. The papers would be lightly reviewed by researchers in another area, simply to evaluate the significance of the development to the wider statistics community. An individual area would only submit a paper after agreeing that it contained a truly new, important result, thus also stemming the overabundance of papers that is threatening to overrun our field. In addition to in-person collaboration at targeted workshops, researchers could post individual investigations on a platform such as arXiv, contact others to further develop ideas, and then only submit to these journals when there is enough collective agreement that a new idea is ready for “canonization.” A newcomer could break in by posting their work on the public platform, thus drawing attention from established researchers and announcing their presence.
With the cost of writing solid research papers rapidly plummeting toward zero, methods for evaluating job applicants and early-career researchers will require major changes. One viewpoint held at present is that strong reference letters from experts are more meaningful than publication counts. Notably, strong evaluation letters could naturally arise for researchers engaging in my proposed publication paradigm. An additional important role of senior researchers who help support this model would be to advocate for junior researchers participating in these collaborations when making hiring or evaluation decisions at their institutions. We might also consider adopting requirements for authors to explicitly specify their roles in developing the work, as is common in the life sciences[9].
What I have laid out may not be a perfect solution, but I have not found another tenable proposal, despite my best efforts. Echoing the thesis of this article, I believe the best solutions are the result of collaborative efforts, and welcome comments, amendments, and suggestions. I hope that this article begins a broader community-wide conversation that I think is absolutely crucial, and also largely overdue.
Acknowledgments
The author is extremely grateful for feedback from Renan Gross, Varun Jog, Laurentiu Marchis, Ankit Pensia, Garvesh Raskutti, Johannes Schmidt-Hieber, Andre Wibisono, and Shuheng Zhou for critical comments on this article, which have led to a more polished and coherent strategy.
Footnotes
[1] See the announcement here: https://openai.com/index/model-disproves-discrete-geometry-conjecture/
[2] The Leiden Declaration (https://leidendeclaration.ai) was released on June 2, 2026.
[3] Perhaps like this Town Hall discussion: https://arxiv.org/pdf/2601.17510 which is now somewhat outdated.
[4] See the article here: https://www.mathopt.org/events-and-discussions/1
[5] The opening paragraphs of Ken Ono’s address may be found here: https://www.linkedin.com/posts/ken-ono-a972191a5_hope-is-not-a-plan-and-nostalgia-is-not-share-7485287590632271872-c6xu/
[6] See Terry Tao’s ICM treatise here: https://arxiv.org/html/2608.16753v1
[7] See the (non-academic) New York Times article on human interactions being replaced with chatbots: https://www.nytimes.com/2026/08/14/magazine/ai-chatbots-internet-communication-loops.html
[8] https://www.nature.com/articles/461879a.pdf and https://dl.acm.org/doi/pdf/10.1145/1978942.1979213 contain interesting insights about the practicality of large-scale collaborative models
[9] See the Contributor Role Taxonomy (CRediT) guidelines at https://credit.niso.org