Math Researchers Admit AI Papers Easier To Generate Than Validate

A team of mathematics researchers at a major university has concluded that it is now faster to generate a publishable proof using large language models than it is to check whether the proof is correct, according to internal emails obtained by Nextish News. The finding emerged during a review of submissions to the Journal of Computational Mathematics, where the editorial board discovered that validating a single AI-generated paper required an average of forty-seven hours of expert human time, while generating a new paper took approximately ninety minutes. The backlog of unreviewed submissions has grown to four thousand eight hundred and sixty-three papers in the past six months alone.

We are now in a position where it takes longer to disprove something than to publish it.

The crisis has created a peculiar incentive structure for researchers seeking tenure and promotion. "We are now in a position where it takes longer to disprove something than to publish it," said Marcus Chen, 51, a mathematics professor at the university who chairs the editorial board. Chen noted that the mathematics community faces a choice between accepting papers it cannot adequately review or rejecting them on principle, knowing that rejection typically delays career advancement for the submitting researcher by one to two years. Most journals have begun splitting the difference, accepting papers provisionally while noting in their metadata that full validation is pending, a status that now describes seventy-three percent of accepted work.

The situation has begun to reshape hiring and promotion decisions across the field. Universities are now recruiting "proof auditors," a newly created position that requires a doctoral degree in mathematics and pays thirty-seven thousand dollars annually, roughly thirty-two percent less than a entry-level software engineering role at a technology company. One major research institution has hired six auditors in the past quarter and is already advertising for twelve more, though the department chair estimates they will need more than fifty to handle the current workload. The auditors work from a shared database of previously flagged errors and known proof-generation artifacts, a document that has grown to two hundred and thirty-four thousand entries and is updated approximately four times per day.