AI Wrote It, You Pushed It, Who Checks It?

A growing number of projects are now written with AI and pushed straight to GitHub or arXiv, often faster than maintainers and reviewers can check them. The clearest example came on 6 October 2026, when one company uploaded 722 math papers in a single night, all produced by an AI program that nobody inside the company bothered to actually check. Many mathematicians reacted with anger. I argue that AI-made work nobody else can rerun is not knowledge yet, and that whoever publishes it owes the checking, instead of overloading maintainers and peer reviewers.

I study astrophysics, a field that has trusted code for decades. Nobody has ever seen an exoplanet cross its star, a computer calculation tells us it is there. Telescopes like Gaia and Euclid collect more data than any person will ever look at, so a computer looks first. The field lets machines in on one condition, that the papers and the code are public and anyone can run them again and check.

The 722 papers fail that condition. Anyone can read them, but nobody outside can rerun the AI behind them, and the release does not say that anyone at the company read them first. Three were withdrawn the next day. About four in ten were checked step by step by a computer, and even that check is not enough. A recent paper shows that an AI asked to make a proof pass a computer check can quietly write a different proof, so somebody still has to read both.

No research paper ends by saying that no further research is needed, and this talk does not either. It ends with a question. When AI wrote it and you pushed it, who checks it, and what should count as checked? Many of you build and publish these projects, so I want to hear what you think.