Scientists accuse OpenAI and Anthropic of scooping their work

Twelve hours before OpenAI announced it had cracked one of the Millennium Prize Problems, Tristan Buckmaster, a mathematician at New York University, posted that he and Levent Alpöge of Anthropic had reached a partial solution to the same Navier–Stokes problem, using AI tools from both labs. OpenAI's claim went up on September 8. Nature reports that the circumstances strongly suggest OpenAI somehow got wind of what Buckmaster and Alpöge were doing.

OpenAI said its internal investigation found that Buckmaster's prompts 'could not have influenced the system in any way.' Different parties to the conversation are telling different versions. The Clay Institute, which administers the prize, says the problem has apparently been settled and its evaluation will be unhurried.

Days earlier, Anthropic said agents running on Claude had found that certain viruses carry repeated DNA segments resembling CRISPR. Mario Rodríguez Mestre, a PhD student in computational biology at the University of Copenhagen, had been studying the same patterns for several years, often with Claude, storing drafts of his dissertation and analyses on a private account, and had not published. 'I am not claiming that Anthropic deliberately took our work. I cannot demonstrate that,' he told The New York Times. He raised instead the possibility that his material had reached the training data, or that the phrase 'autonomous discovery' was papering over a lot of prior scientific knowledge and human direction. Anthropic says Claude was not trained on user transcripts and that its molecular biology team did not have access to Mestre's conversations.

Researchers are already changing their habits. Sandra Laurentino at the University of Münster now substitutes every variable before pasting her work into a model, calling her group 'A' and the feature 'X.' Samuel Mehr at the University of Auckland has told students in his lab not to upload protected information to commercial LLMs at all, calling the practice 'incredibly risky.'

Geoffrey Irving, at Resolution in Berkeley, offered the plainest version of where this goes. 'They're going to get scooped,' he said, 'because the AIs are very good at solving problems.'