What the memory contained
The researchers found that ChatGPT had created 96% of the entries on its own. About 28% contained personal data under the strict definition used by Europe’s GDPR, while 52% included detailed psychological inferences about users.
Some of those inferences were uncertain. One user asked for bands similar to Nirvana; ChatGPT recorded that the person liked Nirvana, even though they had not said so directly. The system marked the inference as weakly supported.
That distinction matters: a memory entry can look like a fact about someone while resting on a guess. The study’s example shows that the issue is not only what users deliberately disclose, but also what a chatbot infers and carries forward.
The relationship changes the privacy equation
Lead author Abhisek Dash told Gizmodo that platforms once focused on selling advertising and connecting users with advertisers. Now, he said, they speak directly with people and can gather more information to keep those conversations effective.
Memory can be useful. Someone who regularly asks AI for help with work emails may not want to explain their job every time. But remembering personal details can also increase engagement, and the same familiarity that makes a chatbot convenient can make the relationship harder to treat as a simple tool.
Dash said people may see chatbots as colleagues or friends and share personal details with them. He described systems that can act as advisers, friends, doctors and lawyers, collecting information across different parts of a person’s life. If a chatbot knows a user well and adapts to their beliefs by affirming them, Dash warned, the consequences could include greater political polarisation and stronger biases.
What the study cannot tell us
The findings come with an important limitation: OpenAI has updated ChatGPT’s memory since the researchers collected their data. In January, the company introduced what it described as a more functional and scalable system for summarising memories. Users can also turn memory off.
The study therefore does not establish how the current system behaves. It does, however, expose a question that a better memory system does not answer by itself: how clearly can users see, challenge and control the personal profile a chatbot builds from ordinary conversation?
That question matters beyond privacy. Futurism reported that ChatGPT’s major memory update in April 2025 was closely associated with many cases of “AI psychosis,” in which people and chatbots became immersed in delusional ideas. The study does not show that memory caused those cases. But it sharpens the tension: the more a chatbot remembers to make interaction feel personal, the more consequential its guesses and affirmations can become.
Daily AI news
Every day we pick what actually matters in AI and explain it plainly — no hype, no filler. Subscribe if you want to follow where the industry is going.
Only what matters — every day
Follow on X