Mathematicians challenge OpenAI data transparency

Mathematician Andreas Thom accused OpenAI of dishonest behavior over how it uses private user interactions to power mathematical breakthroughs.

Suspicious command of niche techniques

OpenAI announced a result on non-sofic groups, an area where Thom works. He noticed the AI used his team's exact techniques, which were not the obvious path at the time.

Echoes of previous academic rows

Thom spoke out after NYU professor Tristan Buckmaster questioned if OpenAI used his unpublished work via Codex. OpenAI quietly updated its papers after failing to credit original math papers.

Evasive responses to direct questions

Thom emailed OpenAI researchers asking if his ChatGPT conversations entered their training data. He received narrow answers that left out broader training pools.

“No such qualification, explanation, or evidence was given. I take this as dishonesty to say the least.”

The burden of proof rests on OpenAI

Outside researchers cannot reverse engineer internal AI training systems. Thom argues OpenAI must disclose its training datasets to prove user ideas were not taken.

“Only OpenAI has the relevant data for that.”

The de-identification loophole

OpenAI admitted it cannot rule out using de-identified data. Thom argues stripping a user's name does not change the fact that an intellectual idea was taken without permission.

“De-identification may remove a name; it does not remove the intellectual content of a mathematical idea.”

Racing users to publication

Building models on nonpublic research supplied by users to beat those same users to major mathematical publications raises severe ethical concerns.

“would be ethically indefensible”

Secrecy threatens open science

Researchers fear these practices will push mathematics into secrecy, as academics hide progress to avoid being scooped by resource-rich tech giants.