Type an unpublished hypothesis into a commercial AI chat window, get useful feedback, and move on. Months later, the same company announces a breakthrough that looks familiar. Two separate research disputes have now pushed that scenario into the real world, exposing a trust problem the AI industry has no clean answer for.
The Disputes
Two researchers say their unpublished work bears an uncomfortable resemblance to results later announced by major AI companies.
OpenAI said an internal model had solved a major mathematical problem. NYU mathematician Tristan Buckmaster raised concerns that the result may have been influenced by related work he had previously entered into OpenAI tools. OpenAI denied directly accessing those chats and said its researchers and agents had not seen the work before it became public.
The allegation remains unresolved. The available reporting does not establish deliberate copying or causal influence.
A parallel dispute surfaced involving Anthropic. University of Copenhagen biologist Mario Rodríguez Mestre said his team had studied the enzyme system Anthropic later described as a novel AI-assisted discovery, and had shared relevant information with Claude. Anthropic said Claude was not trained on user conversations and that its research team had no access to those conversations.
Mestre disputed the implication that Anthropic’s result was entirely independent of prior work in the field. Neither company has been shown to have deliberately taken a researcher’s work, according to the available reporting.
The problem is that the opacity of these systems makes independent verification difficult.
Why Denials Do Not Close the Case
A company can accurately deny training on a specific conversation while leaving several adjacent data-handling questions entirely unanswered.
Retention periods and human or automated access to stored data are separate questions. So are feedback handling, connected tools, and whether information entered operational systems beyond model training. A single denial does not address any of them.
The practical gap shows up in published policy. OpenAI’s terms say content from individual accounts may be used to improve models unless the user opts out; business products and the API are not used for training by default. Submitting feedback on a conversation can allow that conversation to be used for improvement even after a user has opted out of model training.
Opting out, in other words, is not a sealed room: the feedback exception remains open regardless of your settings.
What Researchers Can Do
Enterprise and API arrangements generally offer stronger default protections than personal accounts, but they require active verification rather than passive trust.
Preferring enterprise or API arrangements over personal accounts is a meaningful first step, since those configurations generally carry stronger contractual protections by default. Researchers should still verify retention rules, administrator access, connected applications, and how feedback is handled.
Keeping local records of prompts, uploaded files, model versions, and outputs creates an audit trail that commercial dashboards do not currently provide. For work that is patent-sensitive or involves restricted data, institutional legal or technology-transfer guidance should come before any AI tool does.
The harder problem is structural. Researchers who avoid commercial AI risk falling behind competitors who use it without hesitation. Using it, however, creates unresolved attribution and confidentiality risks that most institutions have not yet addressed through approved tooling or formal policy.
The Bottom Line
The core question for science is not whether to use AI, but whether researchers can use it and still reconstruct how a discovery was made.
Trust will require auditable systems with verifiable records. Company assurances delivered after the fact, however sincerely offered, are not a substitute.
