Mathematician warns against sharing unpublished research ideas with AI tools
Tristan Buckmaster says he still uses AI tools but no longer shares original research ideas with them, citing confidentiality and competition concerns.
Mathematician Tristan Buckmaster says he has not stopped using AI tools, but he no longer shares original, unpublished research ideas with systems operated by OpenAI or Anthropic.
According to the published account, his concern increased after OpenAI released a large collection of mathematical manuscripts generated with an experimental model. The account cites 722 manuscripts spanning 372 families of results. Those figures are not independently documented in the source text and should therefore be treated as claims from the report rather than independently verified totals.
Buckmaster’s broader concern is that young researchers could spend years pursuing a problem while a powerful system produces similar results much faster. Beyond scientific competition, this raises questions about publication timing, intellectual ownership and the protection of early-stage ideas.
He also points to a more immediate risk: reviewers may upload confidential grant proposals to AI services for summarization or assessment. Depending on the service’s policies and account settings, doing so could expose unpublished material to an outside provider.
At the same time, no evidence has been presented that OpenAI took or misused Buckmaster’s work, and the company denies misusing user data. The episode is therefore a warning about handling sensitive information, not proof of a specific violation.
A cautious approach is to review a service’s retention rules, data-use policy, privacy controls and any institutional agreement before entering confidential material. For unpublished ideas or proposals under review, researchers can use organization-approved tools or methods that do not transmit the material to an external service.