Mathematician says OpenAI's 700 AI-written papers hold real ideas but are poorly made

A batch of drafts, and a split verdict

OpenAI released more than 700 math papers on GitHub that were written entirely by AI. They are preprints, meaning early drafts that have not been through peer review, the process where outside experts check a paper before publication.

As technews.tw reported, mathematician Frank Calegari weighed in with a guest post on Terence Tao's blog. He called the release a milestone in the history of mathematics, then listed serious complaints.

What Calegari praises and what he criticizes

Calegari believes the papers show AI can now make breakthrough contributions across several areas of math. Some of the ideas, he says, are excellent.

But the writing is very poor, the papers ignore the field's customs on citing earlier work and giving credit, and some contain errors. He reads this as AI companies not caring much about the mathematics community.

He also tempers the excitement. In his view the papers barely touch some of the central problems in number theory, and solving those would not settle the whole field, just as the periodic table did not settle all of chemistry.

Other mathematicians push back

Ursula Martin of Oxford criticized OpenAI for handing the academic world a pile of messy drafts to sort out. Michael Harris, a number theorist at Columbia, said he will not read them.

An OpenAI spokesperson said the company's aim is to give mathematicians and scientists tools, so it tests and develops its models on hard problems from many scientific fields.

Calegari's advice for the field

He argues mathematicians should focus on what they control: hiring, publishing and training. Jobs, recognition and funding change behavior more than appeals do.

• Assume any research may have used AI help when judging it.
• Interview postdocs and graduate students in person to judge their thinking.
• Have journals raise standards on writing and on importance.
• Teach students a wider range of math while keeping technical training.

He adds that a researcher who picks a worthy problem, gets an AI proof and understands it fully deserves real credit.

Is the game over or just different?

Calegari points to the classification of finite simple groups, after which interest in that area cooled. He says mathematicians should keep exploring rather than treat solved problems as the end of a field.

Pessimists, he says, think the game is over; optimists think it has changed. If AI can prove the problems a mathematician spent a lifetime on, how do we measure that person's contribution?