1. A Max Planck-led study found that ChatGPT created 96% of stored user memories on its own, without any user request.
2. 28% of entries held personal data such as health and location, and 52% captured psychological traits.
3. The study examined the consumer app, not the secured ChatGPT Mil the Pentagon fields for sensitive work.
ChatGPT created 96% of the memories it stored about users on its own, without being asked, according to a February study from researchers at Germany’s Max Planck Institute for Software Systems and partner institutions that analyzed 2,050 memory entries donated by 80 people.
The entries went beyond stated preferences. About 28% contained personal data such as health and location, and 52% captured psychological traits including users’ emotions, beliefs, and goals, the researchers found.
What the Memory System Stored
Researchers obtained ChatGPT histories that 80 users requested under European data-protection law, then identified 2,050 saved memory entries across 65 of them. Only about 4% of those memories followed a direct instruction to remember something. The system wrote the rest.
Most entries, 84%, traced to something a user had typed, which the authors said showed the profiles were largely faithful to the conversations. Others rested on inference. One user asked for bands similar to Nirvana, and ChatGPT recorded that the person liked Nirvana, a preference they never stated.
The team called the resulting profile an algorithmic self-portrait. OpenAI has since expanded the feature. In June, it introduced a system that folds past conversations into an evolving profile of a user’s preferences, goals, and circumstances, and the company says its memory summary does not necessarily show everything the model retains or uses.
A Consumer App the Military Keeps at Arm’s Length
The categories the study flagged, location, health, and psychological state, are the ones security officials treat as sensitive when personnel use commercial AI tools. The Department of War has steered its workforce toward a controlled alternative.
On August 31, it launched ChatGPT Mil on GenAI.mil, a version accredited for controlled unclassified information and built for about 3 million personnel, with inputs kept inside government cloud infrastructure and withheld from training OpenAI’s commercial models. The consumer app the study examined carries none of those protections.
Interim Pentagon guidance has reiterated that operations-security and controlled-information rules still apply to generative AI, and the services have been told to set their own, stricter limits.
On the commercial side, OpenAI began placing ads in ChatGPT earlier this year and says it can use past conversations and memories to target ads when a user turns on personalization, while saying advertisers cannot see conversations or memories. Lead author Abhisek Dash told Gizmodo that a system that serves as an advisor, doctor, and lawyer at once ends up holding information on everything, a concentration he said carries its own risks.
ChatGPT lets users switch off memory, edit the stored summary, or open a Temporary Chat, and separately opt out of model training. OpenAI says deleting a conversation does not necessarily erase a memory drawn from it. The study’s authors proposed an attribution shield, rephrasing questions to strip identifying details, though they said its effect over long chats is unclear.
- arXiv, Dash et al., “The Algorithmic Self-Portrait: Deconstructing Memory in ChatGPT,” Feb. 2026.
- Defense One, Patrick Tucker, “The US military gets its own ChatGPT today,” Aug. 31, 2026.
- DefenseScoop, Brandi Vincent, “New interim DOD guidance ‘delves into the risks’ of generative AI,” Nov. 9, 2023.
- Gizmodo, Ece Yildirim, “AI Chatbots Are Building an ‘Algorithmic Self-Portrait’ of You,” Oct. 22, 2026.
- U.S. Department of War, “Department of War Launches OpenAI’s ChatGPT Mil on GenAI.mil,” Aug. 31, 2026.
Stay Ahead of Global Conflict, Security, and Technology.
Get the #1 daily intel report for decision-makers on global conflict, security, and technology. Over 72,000 leaders trust The SOFX Report to deliver apolitical, fact-based news on issues impacting their lives.
