Infobae flags AI data rights

What you type into a public AI chatbot can be stored, reviewed and, in some products, used to improve the model. Stanford researchers say users often do not realize that. (hai.stanford.edu) (infobae.com) Stanford Human-Centered Artificial Intelligence researcher Jennifer King and her coauthors reviewed 28 policy documents tied to six U.S. chatbot providers: Amazon Nova, Anthropic Claude, Google Gemini, Meta AI, Microsoft Copilot and OpenAI ChatGPT. (hai.stanford.edu) King said the main issue is not one secret clause at one company. It is that privacy rules, retention periods, training defaults and user controls differ across products that people treat like interchangeable chat boxes. (hai.stanford.edu) OpenAI says it collects prompts and uploaded files as user content, and says ChatGPT conversations may be used to train models unless a user opts out or uses Temporary Chat. OpenAI also says business products and its application programming interface are opted out of training by default. (openai.com 1) (openai.com 2) Google’s Gemini Apps Privacy Hub says turning off activity does not stop all processing, and says chats may still be used to respond, protect Google and the public, and support human review. Google separately says Gemini in Google Workspace and Gemini for Google Cloud do not use customer content to train generative AI models. (support.google.com 1) (support.google.com 2) (docs.cloud.google.com) Microsoft draws a similar line between consumer and enterprise products. Microsoft 365 Copilot Chat says prompts and responses are logged for auditing and eDiscovery inside the Microsoft 365 boundary, but are not used to train the underlying foundation models. (learn.microsoft.com) Amazon’s current language for its foundation models says prompts and outputs are not used to train the underlying Amazon models unless the customer consents. Amazon Bedrock also says it does not store or use customer data to train models. (aws.amazon.com 1) (aws.amazon.com 2) Meta’s supplemental privacy policy for Meta Platforms Technologies says information including media and voice recordings used with Meta AI on AI glasses may be used to improve Meta products, including the models that power Meta AI. (meta.com) Anthropic changed its consumer terms in 2025 so Claude conversations are used for training by default unless users opt out, according to Stanford’s summary of the policy change. Anthropic’s commercial terms for Claude on Google Cloud Vertex say customer content is not used to train models. (hai.stanford.edu) (anthropic.com) (anthropic.com) The practical split is now simple: consumer chat products often reserve broader rights over prompts, while business contracts usually promise tighter controls. That makes product choice, account type and logging design part of the privacy decision before a single prompt is sent. (hai.stanford.edu) (openai.com) (learn.microsoft.com) (aws.amazon.com)