Stanford flags AI privacy clauses
- Stanford Human-Centered AI publicized Jennifer King’s review of six chatbot privacy policies, warning that consumer chats can be stored and reused for training. - The Stanford paper examined 28 policy documents and found all six companies appear to train on chat data by default; some may retain it indefinitely. - The study lands as chatbot use spreads and U.S. privacy law remains fragmented. (hai.stanford.edu)
A Stanford Human-Centered AI explainer warned that chats with ChatGPT, Gemini, Claude, Copilot, Meta AI, and Amazon Nova may be collected and reused for training. (hai.stanford.edu) The underlying issue is simple: a chatbot prompt is not just a question, but also data a company can keep, analyze, and feed back into its systems. Stanford researcher Jennifer King said the answer to whether users should worry about privacy is “absolutely yes.” (hai.stanford.edu) King and her co-authors reviewed 28 documents tied to six consumer chatbot products: Amazon Nova, Anthropic Claude, Google Gemini, Meta AI, Microsoft Copilot, and OpenAI ChatGPT. Their paper was posted to arXiv in September 2025 and summarized by Stanford on October 15, 2025. (arxiv.org) (hai.stanford.edu) The paper says all six developers appear to use chat data to train or improve models by default, and some appear to keep that data indefinitely. It also says companies may collect sensitive information disclosed in chats, including health and biometric data, plus files users upload. (arxiv.org) The Stanford team also found that four of the six companies appeared to include children’s chat data in model training. The authors said the policies often omit basic details users would need to understand retention, consent, and downstream use. (arxiv.org) Stanford highlighted one policy change at Anthropic as an example of how fast the ground is shifting. Its October 2025 update said Claude consumer chats would be used for training by default unless users opted out. (anthropic.com) (hai.stanford.edu) OpenAI says it may use content from ChatGPT, Sora, and Operator to improve model performance unless users opt out, while business products such as ChatGPT Team, ChatGPT Enterprise, and the application programming interface are excluded by default. Microsoft says prompts and responses in Microsoft 365 Copilot stay inside the Microsoft 365 service boundary and are not used to train foundation models. (openai.com) (learn.microsoft.com) Google’s Gemini Apps Privacy Hub says turning off activity does not stop all processing, and Google may still use chats to respond, protect users, and create anonymized data to improve services. The page also says human reviewers can help process conversations. (support.google.com) Anthropic’s privacy center now says it uses chats and coding sessions to improve Claude if a user chooses to allow it, while still reserving use in some safety-review cases. That means the exact default can differ by product tier, date, and setting, which is part of the confusion Stanford is flagging. (privacy.claude.com) (anthropic.com) The Stanford paper does not say every chatbot product handles data identically. It says the problem is that privacy disclosures are scattered across policies, subpolicies, and help pages, leaving ordinary users to piece together rules that can change without much notice. (arxiv.org) (hai.stanford.edu) The practical takeaway in Stanford’s write-up is narrower than “never use AI.” Users entering calendars, medical notes, financial details, or private plans are being told to check settings, use opt-outs where available, and think twice before pasting sensitive material into consumer chatbots. (hai.stanford.edu)
