I skipped past the OpenAI marketripe the moment I came to the link in the first paragraph for the Advisory Group on Mathematics and Artificial Intelligence at the Institute for Advanced Study⁠.

Pretty cool that Ed Witten is part of the group 😎

Anyway, check the IAS's Responsible Release of AI-Generated Mathematics -- it establishes some good guidelines usable for other disciplines as well.

Basically:

• cite your fucking references
• write the paper in a style and using terminology that as closely as possible matches existing published norms
• if the paper is not understood by any human yet, provide funding for grad student(s) and/or full on researchers to continue working on it until it is deemed human-understood

TBH it reminds me of the rules our team adopted for agentic coding: it's fine for an AI to write it, but make sure a competent human can easily read it.

Sounds conspiratorial. Which of these problems did you find incomplete or dishonest?

It's mathematics, there's no 'trust me bro' required the proof is either logically consistent or it isn't.

You can check the proofs, they've even provided them in lean: https://github.com/openai/math/tree/main/lean

Downvote is probably because the source is OpenAI so this is just an ad.