SERIES · AI is making papers cheap to produce but they remain costly to referee. Proofs and Prompts is publishing a series of posts from 26 October on the future of publishing, papers and credit. Questions such as how to disclose AI use, who bears the cost of checking, what authorship means and what papers of the future might look like will be addressed. This post is part of the series. Each is written independently and reflects only its authors’ views. All posts in the series
AI disclosure: AI was used for brainstorming and literature exploration.
AI is changing how mathematics is produced. Shouldn’t it also change how we publish?
Do we still need mathematical journals? Spoiler alert: my personal answer is we don’t.
We have known for a long time that our current publishing system is deeply flawed. Before AI saturated every lunch or coffee break discussion, it was one of the recurring topics among (at least French) colleagues. In fact, many have been publicly and vocally advocating for its reinvention for years, see for instance the 2011 post by Timothy Gowers as well as the famous The Cost of Knowledge.
The expectation of a disruption is older still, see for instance this piece of Michael Clarke. Or as a mathematician interviewed by Christian Greiffenhagen later recalled:
“When arXiv was starting to develop, there was a very strong feeling that at some point the whole system would change completely.…Actually it did not.”
Perhaps it is time to make that change happen now that AI has exposed the only remaining purpose of academic journals: produce a metric for hiring and promotions.
What do journals contribute?
Clarke distinguished five functions of journals: dissemination, registration, validation, filtration, and designation.
Dissemination and registration—making results available and establishing priority—were already handled perfectly well by arXiv before AI. Clarke himself acknowledged that journals were no longer required for these tasks.
Validation. This means giving a guarantee that a published proof is correct. We all know that the situation was already highly hypocritical in the pre-AI era. Many reviewers would skim through the introduction before writing a hasty report (most often after having set the paper aside for months to years). Publication in a journal has long ceased to be, if it ever was, a guarantee of correctness. If anything, I trust now much more an LLM audit than a human one, including mine. While I am still on the fence about retiring altogether from refereeing, it is clear to me that I will no longer spend hours tediously checking papers, in particular for journals which deny me the help of LLMs to do so.
Filtration. This means helping mathematicians decide what deserves their attention. I concede that this function might still have some relevance. But do we really choose what to read based on journal publications? In my experience, we learn about new papers mostly through arXiv, discussions with colleagues, and sometimes seminars. By the time a paper appears in a journal, specialists often already know whether it represents significant progress in mathematical understanding or will simply become part of the background noise of mathematical production.
Indeed, the delays between submission and publication are substantial. For instance, the American Mathematical Society’s 2025 backlog report gives median submission-to-acceptance times of 12.2 months for Advances in Mathematics, 14.4 for Analysis & PDE, 16.4 for Algebra & Number Theory, and 17.0 for Annals of Applied Probability. Of course one could object that with the use of AI tools these delays could be significantly shortened. Even still, if AI dramatically shortened publication delays, I doubt journal acceptance would change how we discover or choose what to read.
This leaves designation: publication serving as a proxy for scientific quality in hiring, promotion, and funding. While it has been advocated for a while that it should not be used for that purpose, see in particular the IMU 2014 statement, the CNRS assessment principles, and the San Francisco DORA declaration, regular practice in hiring committees is to rely heavily on these indicators, making designation the main function of mathematical journals. Since everyone seems to agree that the current situation was unsatisfactory before AI and that AI only made it worse (what does it mean now that a paper is at the level of say Annals and not Archive for Rational Mechanics and Analysis?), designation cannot be the reason to save the editorial system. As expressed many times by almost every vocal mathematician, we need to change our evaluation system. I will refrain from trying to discuss how in this post (disclaimer: I have no clue).
To mitigate this last point, there is maybe one positive, although rarely openly acknowledged, role played by journals which is providing a sense of purpose. Indeed, while freedom is probably the most precious part of academic life, it comes with a high price which is total lack of management, clear framework or validation. By this I mean that overall, compared to the vast majority of jobs, we have little to no feedback on our work, notably when we did a good job. Having a paper accepted by a journal, no matter its ‘ranking’, is one of the very few instances of such a validation. We should not underestimate this role as lack of framework might partly explain the high rate of mental health issues among colleagues. Still I would argue that this is not enough to save private Ryan (I am a Millennial after all).
The current system has a financial cost, leaving aside the human cost of managing the journals, and refereeing. According to a survey from COUPERIN, around 78 million euros were spent by public research in France on journal subscriptions (across all fields). Of this amount roughly half goes to Elsevier alone. To put these figures into perspective, there are between 170 000 and 200 000 researchers in France, including PhD students and postdocs. This means that the journal subscriptions would be enough to cover the access of all researchers to the basic plans of either of the main private providers of LLMs with an extra 20 million euros to spare. Of course one could find other and probably better uses for this money.
What shall we do instead?
As for most topics regarding AI the short answer is: I don’t know! A first possibility is to simply rely on repositories such as the arXiv for ‘digested’ papers and Hexagon for ‘raw material’ (see also the recent essays by Frank Calegari or Mathieu Lewin). We could also hope for more ambitious collaborative projects such as We Need ABP or The Bourgain Project. We could also go back to some of the past suggestions from Timothy Gowers and start more interactive platforms where preprints could be discussed and commented on in a MathOverflow-like setting. We could hope that editors decide to voluntarily close journals, or not reopen them. We could also hope that established researchers (I would strongly recommend to non-tenured people to not follow this advice!) would simply stop submitting their papers to journals. As far as I am concerned, I will not forbid my co-authors to submit papers but I will not actively push for it anymore. I had a similar attitude so far regarding submissions to non-academic journals.
The unrealized hopes from the past as well as the inertia of our institutions do not make me very confident that this revolution will indeed happen. Time will tell…
Received 6 November 2026, revised 10 November 2026.

Add to the discussion