On September 7, mathematicians Tristan Buckmaster and Levent Alpöge shared progress on Navier–Stokes after a year of work, including a solution to a simpler version of the problem. According to Nature, OpenAI announced the next day an agent-generated solution. Buckmaster questioned whether research he had uploaded could have influenced the result.
OpenAI denied any such influence. Its investigation found that Buckmaster’s prompts from the previous two months could not have affected the system through training or other means, and the company said neither its researchers nor its agents saw the pair’s work before publication.
The episode points to two concerns for researchers: whether unpublished work stays confidential and whether they can still publish first when AI systems can reach similar findings much faster.
Similar findings do not establish data misuse
A separate dispute concerned CRISPR-like repeated DNA in viruses. After Anthropic announced findings made with Claude, Copenhagen doctoral researcher Mario Rodríguez Mestre said he had studied the same patterns for years with Claude’s help and questioned whether his data had been reused for training. Nature cites Anthropic telling The New York Times that its model was “not trained on any user transcripts.”
So far, no independent verification shows that either company misused researchers’ data. Similar findings cannot reveal how a model reached an answer; establishing misuse would require evidence showing what information the model or researchers could access while the work was underway.
Nor does announcing a discovery confirm that the finding will hold up. As The Neuron’s coverage of the viral DNA findings explains, researchers still do not know the biological function of the reported pattern.
What an upload actually permits
When a researcher submits a prompt, the AI service processes it to produce an answer, a step known as inference. The provider may also retain the conversation, while training is a separate use in which data can influence future versions of a model. An upload alone does not show that the provider retained the material or used it for training.
Under OpenAI’s business policy, business and API inputs and outputs are not used for training by default. Retention has separate controls for qualifying organizations, so researchers handling confidential work need to check their account type and data settings before uploading unpublished material.
Anthropic’s commercial policy says commercial inputs and outputs are not used for training by default. Training may still occur if the user gives permission or submits feedback that includes the conversation. For example, when a user clicks thumbs down and submits feedback, Anthropic may receive the conversation connected to that feedback, not just the single prompt the user originally entered.
For consumer accounts, ChatGPT users can opt out of having new conversations used for training, though the company may still use information users submit through feedback. For Claude consumer accounts, conversations can be used for model improvement with permission, while flagged conversations may support safety enforcement or safeguards training. These policies explain when user data may be used, but they do not show that either company reused unpublished research in the reported disputes.
Precautions can change scientific collaboration
Nature describes researchers already changing how they use these tools. Sandra Laurentino now limits her AI use to debugging and replaces experimental variable names with generic labels. Meanwhile, Samuel Mehr’s laboratory prohibits uploads of protected information to commercial models and discourages their use during research.
Those precautions reduce the unpublished information sent to commercial AI systems, though removing labels can also deprive an assistant of the context it needs to help. Institutions could set clearer rules on which materials researchers may upload to approved AI services and when exceptions are allowed. Vendor assurances should also address retention and access as explicitly as training.
Stricter limits on AI use could also change how researchers collaborate before publication. Researchers who hold preliminary findings until publication give colleagues fewer chances to review methods early, while laboratories with approved private systems can continue using AI during research.
Authorship rules leave priority unresolved
The International Committee of Medical Journal Editors says AI should not receive authorship. Humans remain responsible for accuracy and attribution, and authors should disclose they used AI. These recommendations clarify responsibility for a manuscript, but they do not determine who deserves priority when humans and AI-assisted teams reach similar findings independently.
An AI system could reach a comparable result through public information and its own computational work. Even if no confidential material changed hands, a human team that spent years on the same problem could still lose the chance to publish first. Research records may document how the work developed, but years of work alone do not prove that the team reached the discovery first.
Journals could ask researchers to keep contribution records showing which ideas came from people and which results came from AI systems. Dated experiment logs and preserved model outputs could document how the work progressed, provided confidential records remain accessible only to appropriate reviewers.
Scientific credit involves more than disclosing that AI was used. Journals and research institutions may need standards for determining who originated an idea, who developed the method, and who confirmed the result when human and AI-assisted work converge.
As AI systems become capable of reaching comparable findings faster, those records may become more important for establishing priority. Researchers will need ways to document earlier contributions while keeping unfinished work confidential, especially when publication order no longer tells the full story of who reached a discovery first.
