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
Recently I’ve learned that two separate groups of people have been using AI to make rapid progress on a couple of projects that I had been working on, as well as one problem which is a PhD student’s ongoing thesis project. To be clear, I use AI in research also, and I am not upset that they are exploring what they are interested in. What would bother me is to discover at a late stage that we have both been doing the same things “in private”. And if the student’s thesis came to be considered unacceptable, that would be a disaster.
Much has been said about the idea that we should value exposition over priority of discovery. Maybe. But we are not there yet. And losing the thrill of being the first to discover something, and then sharing it with others, would remove much of the joy of research. Instead, I want to propose something else.
Let’s try being more collaborative within our community. Many people may feel an urge to be more secretive now. I’m not saying to advertise everything you’re doing publicly. But if we have more large working groups within the community, then there is less of a need to rush things out because of real or imagined competition—partly because many of your potential competitors are now collaborators and because you may have more projects going on at one time. There will be less duplication of effort, and collaboration may make our end results better or more impactful.
What would it mean if large collaborations were common? Students may not have to produce solo papers for their PhD theses, but contributing to group collaborations would be sufficient. (This is already the case in many fields.) How would people be evaluated for jobs and grants? Maybe with some combination of publication record and recommendations like today. Maybe there will be an imperfect yet not horrible system of assigning credit for large collaborations. But if we all work with each other more, and know each other more, it should be easier to make evaluations that are less dependent on comparing one person’s publication record with another’s.
I am not saying we should all collaborate in large groups. But however things are in your community now, it will be hard to stay abreast of what others are doing without more effort towards communication and coordination. In fact, I am using the term collaboration rather loosely. I also mean the coordination efforts that many people already do—e.g., multiple small groups/individuals communicating and planning with each other, but writing papers separately.
In the case of one of the problems I mentioned above, I first learned about the situation from a collaborator, who heard about it from one of their collaborators. (In the other case, someone who was working on the problem just emailed me directly). After initially thinking that we may just have to abandon the project, I decided to email the other group to let them know what we had been doing and suggest we coordinate. Then we met on Zoom, and shared what we were doing with each other. It turned out that neither of our works subsumed the other. In our case we decided to continue focusing on the non-overlapping parts separately, and keep each other updated, with the idea that we may join forces at the next stage.
Practically, what do I have in mind, you may wonder. I don’t have in mind public open projects like the Polymath Project or Terry Tao’s recent Siegel zeroes challenge, or massive 200-person collaborations, though those could be part of it. I imagine collaborations that are larger than the typical 2-4 person collaborations—perhaps 5-10, 10-20 or 20-50 person collaborations. Some workshops foster collaborations at these scales, but normally only for a short time, and often these are low-priority side projects for participants. But if you’ve been part of such a project, you may have an idea of what I mean. They involve real human interaction, just like typical small collaborations, but the organization and dynamics may be different.
Large collaborations rarely sprout up organically; real effort and organization are required to start these. Here are some naive thoughts on how to begin:
- Frequent communication between different researchers/groups is key; talk to others as much as you are able—activities like going to conferences, nearby seminars, and seminar meals, or inviting visitors, present lots of opportunities for more communication
- If you are exploring a problem with AI and making rapid progress, don’t race to get out a paper on your own (see also slow maths); contact others who are likely to be interested and see if they want to discuss it and possibly join you
- Organize a workshop or special program that encourages large group discussions and collaborations; from a funding perspective, some could be in-person, some virtual and some hybrid
- It is easier to organize something after forming a core group of people with a shared goal; try to find a few people you trust with common interests and begin from there
- Zulip spaces can be used to help coordinate/organize interest in projects or workshops (e.g., the Proofs and Prompts Zulip space; I also have a dormant Zulip for such purposes in number theory/automorphic forms which could be revived if there is interest)
- I personally favor real collaborations with regular meetings and communication; for larger collaborations, one can break the group into smaller teams for specific tasks; meetings within teams and meetings among team representatives may be more frequent than meetings as one large group; this involves more hierarchy/organization than academic mathematicians are used to in research
- Different collaborations have different dynamics; there’s no one-size-fits-all best practice
- Lots of ideas and attempts at even 2-person collaborations don’t pan out, at least not right away; don’t get discouraged if it takes time to find the right group or project
- If you have other suggestions, please leave them in the comments!
Further thoughts on managing large collaborations
While I am far from an expert in large collaborations, I was encouraged to expand on this aspect. In mathematics, it’s a common view that the time needed to finish a project is inversely proportional to the number of authors. My experience corroborates this, but it also doesn’t have to be this way. First, I think we can look to other fields and industry for best practices for how to manage large projects. Two essays about this from other fields, which I think apply to mathematics with minor changes, are Ten Simple Rules for a Successful Collaboration and Ten simple rules for collaboratively writing a multi-authored paper.
What I have to add to the above 2 pieces is mostly about planning and managing expectations:
- Collaborating with people requires compromise. Understand going in that people won’t go about things the way you think they should (and sometimes that will be worse, but sometimes better).
- Know that both work with AI and large collaborations are new for most of us, and we are all figuring out how to do things; be understanding and forgiving about your missteps and others’. So experiment with strategies and see what people like and don’t like.
- 1-3 people should be in charge of organization and leading the project. These roles require interpersonal and organizational skills; don’t just pick people with the most seniority or mathematical ideas.
- As much as possible, try to set clear expectations and (hoped-for) timelines. For instance, there could be a period of open discussion for 1-2 weeks to come up with goals and plans of attack. Then, either by a democratic or top-down process, decide on a plan, see who wants to work on this plan, and on what aspects. Set working groups with planned inter- and intra-group meetings, and naive goals for when to finish what tasks. Possibly repeat this process at key stages.
- Try to keep goals more concrete and tangible than open-ended. Break-up into manageable subgoals, and require regular meetings/progress reports to keep everyone on track. E.g., in Stage 1, Group A will prove Theorem 1; Group B will prove Theorem 2; Group C will prove and implement Algorithm X. Each group aims to write up their work separately within 3 months. All groups will come together for monthly meetings for status updates and help requests. Then plan Stage 2 together, where everyone combines these results to get Big Theorem.
- Try to set/come to a consensus on expectations as much as possible in advance, with the understanding that you may renegotiate expectations later. In particular, agree on how you will and will not use AI.
- Trust is paramount. Build trust by being transparent about the decision process, and what different groups are doing, and by listening to other people’s opinions.
- Don’t work on another individual or group’s tasks, especially using AI, without asking them first.
- Make contingency plans for when it inevitably happens that, for whatever reason (health, other responsibilities, …), someone cannot contribute or complete their tasks on a satisfactory timeline. For example, there could be an advance agreement that the project will go on without indefinitely waiting on people, and they will stay as co-authors and contribute what they can unless they opt out of the project.
- Relatedly, upon review stages of the work, set time limits for feedback and comments.
- Understand that everyone will not contribute “equally”, and that this is fine. If desired, a contributions statement could be added. Be generous about acknowledging others’ contributions and communicate openly to make sure you are all on the same page.
Thanks to John Voight, Sachi Hashimoto, Roi Docampo, and Proofs and Prompts hosts for some feedback on drafts of this post. AI tools were used for light proofreading and searching for relevant links.
Received 6 November 2026, revised 21 November 2026.

Add to the discussion