NEW YORK—The editorial board of the Quarterly Review of Rigorous Mathematics announced this week that it will send each of the 53 results from Google DeepMind's AlphaProof Nexus to three anonymous human referees, and expects to finish "by 2034." The decision comes days after journals expressed concern about proofs arriving already checked.

"Just because a proof has been formally checked by a computer, line by line, with no room for error, doesn't mean it has been reviewed," said managing editor Martin Brenning. "Review means a tired associate professor reads it on a plane."

The paper, published Thursday in Science, reported that the system attempted 353 Erdős problems and solved nine, two of which had been open for more than 50 years. It also settled 44 of 492 open conjectures from the On-Line Encyclopedia of Integer Sequences. Every proof was checked in Lean, at a cost of a few hundred dollars per problem.

According to journal records, the Quarterly currently takes a median of 31 months to return a first-round referee report. Brenning said that number "reflects our standards."

Invitations have gone out. Of the first 40 mathematicians asked, 38 declined, one accepted by mistake, and one asked whether the journal could just run the Lean files again. That request was escalated to the board and denied.

"If software checks the proofs, then what exactly are referees for?" Brenning asked, and then took 11 seconds to remember.

Associate editor Carol Weintraub said the scrutiny is about fairness. "Every human author waits three years and gets a report that says 'see attached,' with nothing attached," she said. "It would be wrong to hold a machine to a lower standard of suffering."

The board also approved an article processing fee of $3,400 per proof, which Weintraub called "in line with industry norms" and "only somewhat more than it cost to produce the mathematics."

At press time, the first referee report had arrived early, in 2031, requesting that the authors motivate the result more clearly.