ChatGPT may know more about you than you can actually realise: How AI memory raises questions about privacy

ChatGPT may retain more about its users than they consciously choose to share or ask it to remember, raising questions about privacy, consent and the growing role of artificial intelligence in storing personal context.

A study titled The Algorithmic Self-Portrait: Deconstructing Memory in ChatGPT examined 2,050 memory entries donated by 80 users. Researchers analysed who initiated the creation of these memories, what information they contained and whether the entries reflected details users had shared in their conversations.

The findings suggest that ChatGPT’s memory feature relies heavily on the system’s own decisions about what information to retain. However, the researchers cautioned that the results reflect their dataset, which was relatively small and comprised volunteers who agreed to share their data.

ChatGPT creates most memories without explicit requests

Around 96% of the memory entries analysed in the study were created by ChatGPT without users explicitly asking the chatbot to remember the information. Only about 4% followed direct requests from users.

This distinction matters as AI assistants increasingly use information from previous interactions to make responses more relevant and personalised. Details shared while seeking help with work, relationships or everyday decisions may remain available beyond the conversation in which they were disclosed.

The study found that approximately 84% of the entries could be traced directly to something users had said. This suggests that the system was largely retaining information drawn from conversations, rather than generating entirely new details about individuals.

Personal details and psychological insights in AI memory

The researchers identified two significant categories of information in the stored entries. About 28% contained personal data, including details relating to health, relationships and location. Another 52% contained what the researchers classified as psychological insights, covering aspects such as personality traits, motivations, goals and states of mind.

These findings do not establish that ChatGPT secretly monitors users’ typing speed, hesitation or emotional states. The study examined the content of saved memory entries; it did not demonstrate that the chatbot maintains a hidden, continuously updated psychological assessment of every user.

Nevertheless, retaining personal context can have implications for privacy. Information shared to obtain a more useful answer may influence subsequent interactions, even when the user does not expect it to persist.

What does this mean for user privacy?

The findings highlight a broader question for AI companies: how should systems balance personalisation with meaningful user consent?

A notification that information has been saved may not necessarily mean users understand what has been retained or how it could shape future responses. Greater transparency over memory creation, alongside clearer controls over stored information, could help users make more informed choices.

Users can review ChatGPT’s memory and personalisation settings, remove saved memories and adjust whether the service refers to saved memories or past chats. Deleting a conversation alone may not remove a memory created from it, so users should check both where applicable.

The study does not prove that ChatGPT knows users better than they know themselves. It does, however, draw attention to a less obvious consequence of AI personalisation: information shared in one conversation can become context carried into another, making the boundaries of consent and data retention increasingly important.