Scientists rethink AI use over fears of being scooped

Some scientists are restricting their use of commercial AI tools after companies announced discoveries overlapping with their unpublished research. The disputes highlight concerns about intellectual property and whether increasingly capable AI agents could independently beat researchers to important results.

Mathematician Tristan Buckmaster questioned the timing of OpenAI’s announcement that its agents had solved a simplified version of the Navier-Stokes problem. Buckmaster and collaborator Levent Alpöge had publicly shared their own solution shortly beforehand. OpenAI denied that their private prompts influenced its system and said its researchers and agents had not seen the collaborators’ work before its public release.

A similar controversy followed Anthropic’s announcement that Claude-powered agents had identified repeated DNA patterns in certain viruses resembling those in CRISPR systems. Computational biology researcher Mario Rodríguez Mestre said he had been studying the same patterns, often using Claude, without publishing his findings. Anthropic said its model was not trained on user transcripts.

The reassurances have not restored trust for everyone. Epigeneticist Sandra Laurentino now limits AI use to debugging code and disguises experimental details. Cognitive scientist Samuel Mehr has barred his students from uploading protected information to commercial models and warns against their use throughout research. Geoffrey Irving, chief scientist at AI-safety organization Resolution, argues that researchers could increasingly be scooped because agents become better problem-solvers, rather than because private manuscripts enter training data.