The purpose of this letter is to warn the mathematical community of a new type of paper mill that could potentially appear as a consequence of the advances in artificial intelligence. I am a postdoc working on Math and AI, and I organize an international online seminar on math and machine learning.
Firstly, there exist groups that produce fake research papers and sell them to scientists; these groups are commonly called paper mills. According to this study, at least 2% of all research submissions in all sciences are submitted by paper mills.
Secondly, there also exist start-ups, companies and universities working on solving mathematical problems with machine learning. In particular, some use tools that take information from a mathematical problem and return, when possible, experiments to prove it. The level of human interaction varies: in some cases the user has to provide only a text description of the problem; while for other programs, the user must write code that evaluates the fitness of the solutions proposed by agents.1 AI agents are also capable of transforming the output of those experiments into a written research paper. While it is not clear how much interaction was needed to produce the preprint about the Navier-Stokes problem or to solve the Jacobian conjecture, in this post I want to give emphasis on the possibility of an artificial intelligence solving completely a problem only using a prompt as input.
With these two antecedents, I would like to ask the reader: what would happen if a group trained an algorithm to recognize important mathematical questions? The algorithm can work in parallel and mass produce meaningful questions. If we now give this list of questions as inputs to algorithms that try to solve them, and we furthermore assume that they manage to solve a small percentage of the input questions, then the final products are research papers that are publishable since they have solved meaningful questions for mathematicians. Finally, what if the company sells those finished results to mathematicians? This would create a paper mill that produces correct papers and sells the authorship. Moreover, if an AI that solves problems has a certain level of quality, an AI that formalizes the proof can further filter the papers.
One can wonder what the danger is if paper mills are producing knowledge, that is, publishable papers. Fraudulent credit is one of the problems; affecting hiring, promotion, grants, and reducing the credibility of the publication process.
In informal conversations, I asked some of the experts who attended the Mathematics × AI: Challenges and Opportunities event at the Korea Institute for Advanced Study and the International Conference on Machine Learning 2026 about the possibility of this new type of paper mill. Some of the people I talked to are collaborating with AI companies, while others were participants with different levels of experience in AI.
While I personally expected to be assured of the impossibility of this new type of paper mill, some of the experts considered this scenario plausible within a few years, and they could not think of a solution for journals to detect this problem. Some told me they have students working on an algorithm to predict important problems. I was encouraged by them to keep talking about this problem in order to raise awareness, which led me to write this letter.
Why would this be a very difficult problem for journals?
From the point of view of the journals, there are no steps in the review process to ensure that the person who submitted the paper did not buy the paper from a paper mill, or to certify that the person is aware of the content of the paper. Note also that using a detector to see if a paper is written by AI is useless: a person can buy finished papers and rewrite them.
We are now seeing talks in which people used AI in their work. Note that the quality of the talk may not be related to the use of AI. The speaker could be afraid of public speaking, perhaps the speaker does not have enough experience giving talks, faces a language barrier, or is new to the topic and knows just the minimum required for the result to hold.
I call on the community to discuss what changes journals would need to implement if we ever reach that level of technology.
Acknowledgement: I would like to thank Susana Lopez Moreno for related discussions.
- The following examples, listed alphabetically, are not exhaustive: Aletheia, AlphaEvolve, Astra, Iteris, OpenEvolve, Rethlas, and ShinkaEvolve. ↩︎
Received 11 September 2026.
Add to the discussion