Scientists Restrict OpenAI and Claude Over Privacy Fears

On September 7, mathematicians Tristan Buckmaster and Levent Alpöge shared a year of progress on the Navier-Stokes equations. The very next day, OpenAI announced its own agent-generated solution to the problem, prompting Buckmaster to question if his previously uploaded prompts had influenced the model. OpenAI denied any influence, stating that Buckmaster's prompts from the prior two months could not have trained the system. Similarly, Copenhagen researcher Mario Rodríguez Mestre questioned whether Anthropic used his data to find CRISPR-like viral DNA patterns using Claude, though Anthropic stated the model was not trained on user transcripts.

While there is no independent verification of data misuse, these incidents highlight growing anxieties over intellectual property. Both OpenAI and Anthropic state that their commercial and API inputs are not used for model training by default. However, consumer accounts and user feedback mechanisms, such as clicking a thumbs-down button on Claude, can still result in conversational data being processed for model improvement.

To mitigate these risks, scientific practitioners are actively changing how they interact with AI. Some researchers, like Sandra Laurentino, now restrict AI use to debugging code while replacing experimental variable names with generic labels. Others, like Samuel Mehr, have instituted laboratory policies that outright prohibit uploading protected information to commercial models.

Beyond data privacy, the speed of AI development threatens the traditional academic reward system. Even if an AI system independently reaches a breakthrough using public data, human teams who spent years on the same problem risk losing their chance to publish first. Because the International Committee of Medical Journal Editors rules out AI authorship, institutions must now establish new standards to document human contributions and determine scientific priority when human and machine discoveries converge.