Researchers question OpenAI's reliability with unpublished math
openai
| Source: HN | Original article
Mathematicians are increasingly questioning whether OpenAI can be trusted with unpublished research, after multiple scholars allege the company has used their work without permission.
OpenAI’s claim of solving the Navier–Stokes Millennium Prize problem has been eclipsed by fresh accusations that the company may have drawn on unpublished mathematical work. A second mathematician, Andreas Thom, has publicly alleged that OpenAI gave a misleading answer when asked whether his private research, shared via the company’s Codex platform, was incorporated into its training data. The complaint follows a similar charge made days earlier by Tristan Buckmaster, which we reported on 10 September 2026.
The dispute centers on Codex, OpenAI’s coding and research assistant, where researchers routinely discuss and develop cutting‑edge proofs. Thom argues that the only party capable of confirming whether his unpublished results were ingested into the model’s training set is OpenAI itself, and that the firm has not provided the necessary evidence. The lack of transparency raises a broader trust issue for scientists who rely on AI tools: can they safely explore frontier problems without risking inadvertent exposure of their work?
If the allegations prove accurate, they could undermine confidence in AI‑assisted research and prompt calls for stricter data‑use policies. Stakeholders are likely to watch for an official response from OpenAI, potential third‑party audits of its training pipelines, and any regulatory scrutiny that may arise. The episode also puts pressure on the AI community to establish clearer safeguards for unpublished research, a step that could shape how future breakthroughs are pursued in collaboration with powerful language models.
Sources
Back to AIPULSEN