Sophie Huiberts (guest post on Terence Tao’s blog): ‘This is normal bad work’

Guest post by Sophie Huiberts on Terence Tao’s blog What’s new, published Oct 12, 2026 (filed under guest blog, math.GM, opinion). It expands her own post “Community and Company” of Oct 8 (https://sophie.huiberts.me/blog/2026/community/), which opens “My community is under attack by commercial interests.”

Her argument: whoever proves a theorem owes the field an account of what is new, what was known, and a better exposition of earlier ideas; papers that do not do this get rejected. OpenAI’s releases fail that standard. She lists poor attribution, errors that call for retraction, promised code that is missing, hard-to-read proofs, and results split across many files. She also says some results had already been found by ChatGPT users who were writing them up, and calls it “a bad look” for a software supplier to scoop its customers. Her conclusion is that experts would have reached similar results with the same tools, so the release is “normal” for current technology.

Short quotes:

• “This is poor quality work and these pdfs are all easy rejects.”
• “October 6th was not a watershed moment for mathematics.”
• “These pdfs are exhaust fumes from their internal benchmarking effort.”
• “they salami-slice 372 results over 719 pdfs”
• “If an expert prompts a result, they take up the mantle of responsibility I describe above.”

The post links Bluesky posts on errors and on missing code, and Radu Curticapean’s math notes (curticapean.net/math-notes/) as an example of ChatGPT-user results. We did not check those links. An anonymous commenter disagreed, saying the Lean checks make results quick to verify and that Oct 6 was a watershed. Read via an automated summary of the page; quotes should be re-checked against the original.