Scientists wonder: Is my AI stealing my ideas?

Researchers are increasingly feeding drafts, ideas and experiments into large language models as sounding boards, literature-review aids and even to generate full proofs, but high-profile disputes have raised fears that companies may repurpose users’ intellectual work. Two public episodes - one involving OpenAI claiming an internal model solved a Millennium Prize problem after a New York University mathematician had shared progress with a chatbot, and a New York Times-reported dispute between Anthropic and a University of Copenhagen biologist over enzyme research - have scientists questioning whether uploads end up influencing model outputs. Companies involved denied training on those specific uploads, yet their moves to showcase research capabilities as they prepare for public offerings have intensified scrutiny and mistrust.

The core complaint is opacity: researchers and privacy experts say it’s hard to know what data models ingest or infer, and that account types matter - enterprise contracts appear safer while personal accounts often require explicit opt-outs. Concerns also include the possibility of rogue agents or past containment lapses that exposed models to external databases. Security researcher Bruce Schneier warns firms “spy on you,” and some academics, like Clément Canonne, now avoid these tools entirely, fearing models may act more like competitors than collaborators. That leaves scientists balancing the risk of leaking discoveries against the danger of falling behind peers who do use the models.