What does AI actually know about you?

Summary

People often type sensitive information into artificial intelligence tools that they would never broadcast publicly, ranging from medical questions and money worries to drafts of nervous emails. This creates a natural question regarding what an artificial intelligence actually knows about a user. Zoe Ludwig from the education team at Anthropic walks through what happens when individuals share information with AI, noting that policies vary between providers and users should always confirm specific tool details. The core principle is that users should be able to see and control what their AI provider retains, focusing on where data goes, why it goes there, and how long it stays.

Data shared with AI moves through three primary channels, starting with the conversation itself. During an active chat, the model holds context for the duration of that specific conversation, such as remembering a user's mention of a partner, a love of hiking, and proximity to Denver when recommending local activities. Once that chat is closed and a fresh one is opened, the model starts completely blank and has no recollection of those details. Control in this scenario is straightforward because the model only knows what the user puts in front of it.

The second channel involves product memory, where many AI tools save details from one chat to the next. This storage happens in the user account rather than the model itself, allowing the product to reference preferences in future interactions, such as factoring in a love of hiking into weekend plans. Depending on the product, this memory feature may be enabled by default or optional, and users maintain the ability to edit, clear, or turn off memory entirely. The third channel involves the provider's own systems. Privacy policies often allow data use for running services, safety and abuse review, bug fixing, research, commercial purposes like serving ads, and model improvement. Closing a chat keeps the information with the provider, making it essential for users to check privacy policies and data retention settings.

Some providers may also use conversations to train future versions of their models. In these cases, personal data is typically removed from chats before training, and conversations are not stored in the model to be read back to other users. However, once a chat is closed, the details exist as patterns rather than exact words, and many providers allow users to opt out of this training. Practical habits can help users maintain control over their information. Users should spend a few minutes reviewing settings for memory, chat history, connected apps, and training toggles. Sharing should be based on personal comfort rather than a blanket rule, and sensitive personal details can be omitted or replaced with placeholders like email recipients. Finally, users should match the product to data sensitivity, using consumer versions for everyday tasks and business or enterprise plans with specialized data handling terms for confidential or regulated work.