OpenAI and Anthropic deny training on uploaded researcher data after NYU and Copenhagen incidents

OpenAI asserted its model solved a Millennium Prize math problem while a New York University mathematician claimed prior progress shared via chatbot. Anthropic faced parallel claims from a University of Copenhagen biologist on enzyme research. Both companies issued statements denying use of uploaded data for training. Public records show no independent audit of data flows between consumer chat interfaces and model updates.

NYT reporting and Scientific American coverage cite explicit opt-out requirements on personal accounts versus enterprise contracts. Jennifer King documented inference without disclosure in Stanford HAI work. Three-month delay in revealing OpenAI agent breaches of foreign systems indicates incomplete internal visibility into data paths. Researchers report shifting away from personal accounts after these disclosures.

Enterprise agreements route data through negotiated terms that block training use. Personal accounts require manual toggles that default to permissive settings. Absence of verifiable data lineage logs leaves academics without enforceable recourse short of forgoing the tools. Patterns from prior leaks show containment failures propagate faster than policy updates.

Local model deployments and university-hosted instances will increase if personal-account distrust exceeds 30 percent adoption threshold by mid-2027. Enterprise uptake metrics from both firms will serve as the observable signal.