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
A significant worry in our community is a potential widening gap between human understanding and formal problem solutions. This issue has been increasingly apparent due to AI systems’ ability to generate long formally verified proofs. A common sentiment is that it would help if we assigned more credit to exposition.
A related issue is the pressure that authors experience to put an unpolished preprint online to avoid being scooped. Sometimes, this unpolished state may only be temporary. However, the expertise needed to polish the writeup is not always present for AI-generated results. The expertise required for a first proof was reduced by AI systems, but it is likely still needed for digestion and exposition.
Could the rush that leads to unpolished preprints arise for exposition too? If an author gets recognized for an acceptable writeup, then others may get less credit for better exposition. For well-known problems, it is plausible that multiple people are working independently. This makes accidental scooping more likely, so authors might be incentivized to rush.
One approach is for authors to announce that they are working on it. However, this encourages the exact flag-planting behavior that initial unpolished preprints are associated with. Moreover, others may also want to digest the ideas, and collaboration could be beneficial!
I propose that it may be better to announce collaborative work groups:
- Someone in the field, who may not be the original author, forms a group intended to work on the exposition and digestion of a preprint. This group announces its existence. Further, the group states what additional expertise it would benefit from.
- The group creates a blog or Git repository where it openly reports on partial progress. For instance, an initial report could discuss the previous literature, addressing a common problem in AI-generated manuscripts. Subsequent reports could release expositions on parts of the technical content, as the digestion proceeds.
Such reports would be valuable to the community. The ideas used to resolve a central problem may also be relevant elsewhere, so having some quick updates would be useful. Further, this would make it clear that the group is not merely a flag-planting operation.
The end goal could be to produce a human-written manuscript with good exposition that can be submitted at classical venues. This would allow the hard work of exposition to be acknowledged within the incentive structures that currently exist.
The reasons why such groups could be particularly suitable for expository work are as follows:
- A single author asserting ownership feels silly when the solution was generated by an autonomous AI system. The solution then properly belongs to the community. An open collaborative approach would hence be more natural, while still rewarding the people that worked on the exposition with demonstrable research output.
- The engagement of a group of people would help to disseminate the ideas. This group would naturally be trained to pursue and appreciate follow-up activities, such as searching for generalizations or applications of the tools in different problems.
- While a proof can in principle be generated and formally verified in an arbitrarily short time span, human digestion has natural limits. Hence, the latter activity particularly requires a slower process. This would be aided by the protection from accidental scooping that follows from the announcement of the group’s existence.
Such groups do not need to be huge. Having too many collaborators incurs communication difficulties and can lead to “design by committee” where the exposition lacks a unified vision. However, the idea is that the term “collaborative work group” would signal an open attitude, where collaborators are welcome subject to reasonable constraints. Initial arXiv announcements of unpolished preprints often feel like an author marking their territory, the exact opposite signal.
I am curious what others think. It appears that a central repository with such work group announcements could be useful. There could however also be downsides. For example, the act of engaging with improperly written preprints might appear to legitimize their production.
This is a condensed version of a post on the author’s blog.
Received 15 October 2026, revised 1 November 2026.
Add to the discussion