As AI agents become increasingly capable of conducting scientific work autonomously, researchers are facing questions that once seemed unthinkable: could an AI system solve a problem or publish results before the scientists who have been working on it?

At least twice in the past five weeks, researchers said they had spent years working on a research question only to discover that an artificial-intelligence company had answered it or published results before they were ready to do so. At the same time, some scientists are restricting their use of AI tools because they fear that unpublished research could be exposed or used to improve commercial models.

Anthropic’s AI Biolab has discovered “CRISPR-like” DNA in viruses. What’s next?

Some, although not all, AI researchers say it is becoming increasingly likely that AI agents will reach scientific conclusions before human researchers do.

“Researchers are getting scooped, but not because they uploaded a manuscript,” says Jeffrey Irving, chief scientist at Resolution, an AI-safety research organization in Berkeley, California. “They’re going to get the scoop because AI is very good at solving problems. And AI is going to get better and better.”

AI research sparks controversy over the Navier–Stokes problem

On 7 September, Tristan Buckmaster, a mathematician at New York University in New York City, posted online that he and his collaborators had made progress on the Navier–Stokes problem, a long-standing unsolved problem in fluid mechanics.

According to Buckmaster’s statement, he and Levent Alpöge, a mathematician at the San Francisco-based AI company Anthropic, had been pursuing the problem for a year as a personal collaboration. Buckmaster’s post included a paper offering a solution to a simpler version of the problem.

The following day, 8 September, OpenAI, also based in San Francisco, announced that one of its AI agents had solved the puzzle. Buckmaster publicly questioned the timing of the announcement.

Who will be valued in the age of AI? OpenAI’s mathematical breakthrough sparks debate

Buckmaster raised the possibility that information he and Alpöge had uploaded to an OpenAI tool could have been used to train the company’s models. OpenAI disputed this scenario.

In an investigation statement, OpenAI said that prompts submitted by Buckmaster during the two months before the 8 September announcement “could not have influenced the system in any way, including through training”. The company also said that its researchers and agents “didn’t see any of their information”.

Scientists question whether AI tools can protect unpublished research

Two weeks later, Anthropic announced that an agent running on its Claude large language model had discovered that some viruses contain repetitive DNA segments resembling patterns found in the CRISPR gene-editing system.

After the announcement, Mario Rodríguez Mestre, a PhD student in computational biology at the University of Copenhagen, told The New York Times that he had been studying the same DNA pattern for several years.

“I often use Claude, but I don’t publish any of my work,” Mestre said. He also raised the possibility that information his team had uploaded to Anthropic’s tools could have been incorporated into Claude’s training data. Anthropic said that the model is “not trained on user transcripts”.

Scientists are becoming more cautious about commercial AI

Despite these reassurances, some scientists are reevaluating how they use AI systems.

Sandra Laurentino, a reproductive epigenetics researcher at the University of Münster in Germany, says she no longer trusts AI systems with details of her research. She uses AI only to investigate why her code is failing, and then changes all parameter and variable names to generic labels, such as “Group A has feature X”, so that the system receives no information about the experiment.

“I’m pretty cautious when it comes to AI,” she says, adding that recent controversies have “made me even more paranoid”.

Samuel Mair, a hearing cognitive scientist at the University of Auckland in New Zealand, has also become more hesitant to use commercial AI models. He created an AI-usage policy for his laboratory that prohibits students from uploading protected information to commercial large language models and warns them not to use the systems at any stage of the research process.

“I think it’s very risky to hand over your intellectual property to a third party when you don’t know what they’ll do with it,” he says.