ChatGPT may be building a picture of users' personalities, preferences and mental states by turning everyday conversations into stored memories, a study has warned. Researchers found that many memory entries were initiated by the AI system without an explicit request from the user, raising questions about privacy, transparency and control over personal information.
The study, The Algorithmic Self-Portrait: Deconstructing Memory in ChatGPT, examined 2,050 memory entries drawn from conversations donated by 80 ChatGPT users. The findings concern stored memory entries in the study's dataset, not comprehensive psychological profiles of every ChatGPT user.
ChatGPT Initiated 96% of Recorded Memory Entries
Researchers found that 96% of the entries were created unilaterally by ChatGPT, with only 4% following an explicit user request to perform a memory operation. The 2,050 entries came from 1,058 conversations involving 65 of the 80 participants.
The researchers' classification means no explicit memory-operation request was detected; it does not establish that users lacked access to memory controls or that OpenAI breached its policies.
Researchers found ChatGPT is secretly writing a psychological profile of every user.
— Superman (@thesupermannx) October 20, 2026
They published a paper that exposes what ChatGPT is really doing to you.
And the results are terrifying.
96% of the things it remembers about you were stored without you ever asking.… pic.twitter.com/SzLK5G0iXz
The findings raise questions about how much users understand about what an AI assistant retains from conversations that may initially appear casual or task-specific.
Lead author Abhisek Dash told Gizmodo: 'Earlier, the platforms had one objective: to sell you ads and to mediate an interaction between you and an advertiser. But now, they are actually talking to you, so they may want to know more about you so that they can have an effective conversation.'
Personal Data and Psychological Insights
The study found that 52% of the memory entries contained at least one psychological category, such as a user's desires, intentions, emotions or beliefs. Meanwhile, 28% contained personal data classified under the EU's General Data Protection Regulation (GDPR), and 7% contained special-category data.
These figures refer to memory entries, not the percentage of users affected, and the categories can overlap. The GDPR classifications reflect the researchers' analysis, not a legal finding that OpenAI violated data-protection law.
One example highlighted by the researchers involved a user asking ChatGPT for bands similar to Nirvana. The system recorded that the user 'likes Nirvana', although the request did not explicitly state that preference. The researchers described the entry as an assumption and rated it as weakly supported.
The example illustrates how a system may turn a request into a broader statement about a person. However, the researchers also found that 84% of entries showed direct textual overlap with users' wider conversation histories. That overlap suggests many memories were grounded in the conversations, but it does not prove that every entry was accurate or free from inference.
Dash also warned of wider risks, telling Gizmodo: 'That is the privacy risk which is sort of emerging, because there is no boundary anymore; the system is acting as your advisor, your friend, your doctor, your lawyer, and so it has information on everything.'
He raised concerns that chatbots could reinforce users' existing beliefs and contribute to political polarisation. These are potential risks identified by the author, not outcomes demonstrated by the memory study.
How to Review ChatGPT's Memory
OpenAI says memory helps ChatGPT personalise responses by retaining useful context from conversations. On 4 June 2026, the company announced a rollout of an updated memory system, initially for Plus and Pro users in the US, with wider availability planned.
Users can review available controls under Settings > Personalization > Memory, although options vary by plan, region, platform and workspace. Turning off memory does not delete past chats. Removing remembered information may also require deleting the original conversation and other sources where that information appears.
The study highlights the tension between useful personalisation and privacy as AI assistants retain more context across conversations. Its findings do not establish that ChatGPT routinely invents characteristics about users, but they raise questions about how inferences are made, stored and communicated.
