What happened
Scientific American’s Peter Hall describes scientists’ growing concern about the privacy of manuscripts and ideas they share with AI tools. Researchers use large language models in varied ways: as sounding boards, for literature review, to clean up nearly finished work, and, increasingly, to generate proofs. The article sets those uses against two high-profile disputes in which companies’ public research announcements followed researchers’ use of their products.
In the first, OpenAI said an internal model had solved one of the seven “million-dollar” math problems. A mathematician at New York University insinuated that the company had taken his progress, which he had shared with one of its chatbots. Weeks later, The New York Times reported a similar story involving Anthropic and a biologist at the University of Copenhagen. The biologist had uploaded unpublished work concerning a group of enzymes. Hall says both OpenAI and Anthropic stated unequivocally that they had not trained their models on the respective users’ uploaded data.
The article does not establish that either company took or used the researchers’ ideas. It says there is no evidence that this happened in either case. Rather, it reports how the episodes have affected researchers’ trust. Some scientists continue to use these systems, while others are starting to see them as competitors. Clément Canonne, a theoretical computer scientist at the University of Sydney, says he is not using the tools and is not encouraging their use. He also describes a pressure to keep up: researchers may worry that if they avoid the models, they will become irrelevant or lose ground to people who use them.
Jennifer King, a privacy and policy researcher at Stanford University’s Institute for Human-Centered Artificial Intelligence, says users should examine their account settings. According to King, personal accounts may require users to opt out explicitly to prevent their information from being used for training. She describes enterprise accounts as seeming safer, while warning that transparency about company practices is limited. Hall also reports that some researchers worry about what agents might access, but the article does not show that an agent accessed the disputed material. It gives no pricing, account-by-account policy comparison, or specific settings instructions.
What the archive adds
We have not covered this before; it opens a new thread.
Trend signal
This does not yet fit a pattern New Runtime tracks. The article describes a trust problem around researchers’ data and unclear practices, but it does not document a shipped security boundary, a permissioned capability, or a technical control that would substantiate one of the patterns New Runtime follows.
Why now
Scientific American places the disputes amid AI companies’ efforts to demonstrate their models’ research capabilities. The timing matters because researchers may feel pressure to use more capable systems even as the work they submit becomes more valuable and sensitive. The article also refers to reported rogue events over the summer as a factor in some users’ heightened concern; it does not connect those events to the researchers’ specific uploads.
Why it matters
For teams building with AI, the practical issue is not whether a chatbot is useful; researchers are already using models for tasks ranging from literature review to generating proofs. The issue is whether the team can confidently decide what work is safe to send, under which account, and under what data-use settings. If that decision depends on assumptions about training, retention, or agent access that users cannot verify, sensitive workflows may be blocked regardless of model capability.
The article also shows why a company’s denial of training on a particular upload may not resolve broader concerns. Researchers are worried both about deliberate use of their content and about limited visibility into how systems handle it. For product and research teams, this points to a need for explicit internal rules: classify unpublished material, check the relevant service terms and account settings, and avoid treating an “enterprise” label as a complete security guarantee. Those are sensible precautions, not proof that any particular service mishandled data.
Limits and uncertainty
The article does not independently establish what happened to the mathematician’s or biologist’s uploaded work, and it explicitly reports no evidence that either company’s research announcement resulted from taking the users’ ideas. OpenAI and Anthropic deny training on the respective uploads. The article does not measure how many researchers have stopped using AI, or quantify the prevalence of the concerns it describes. It also does not provide a technical audit of data retention, agent permissions, or account controls across providers. King’s comments about personal and enterprise accounts are expert guidance, not a comparative verification of every product’s current settings.
Reader takeaway
The article documents a real trust barrier, not proof of idea theft. If your team puts unpublished research or other sensitive work into an AI service, check the account’s data-use settings and your organization’s policy before uploading. Treat enterprise status as a factor to investigate, not as a substitute for understanding what the service can access and how it handles your data.