Researchers presenting their findings at the ACM Web Conference 2026 analyzed EU user chat histories and discovered 2,050 entries belonging to 65 individuals stored within ChatGPT's memory. According to reports, the system retained the lion's share of this information without any direct prompt or explicit user consent. In effect, the chatbot accumulates a detailed digital profile of a person—ranging from habits and interests to behavioral patterns—and uses it to personalize responses.
Algorithms Invent What Never Existed
The study's authors paid special attention to ChatGPT's ability to make its own assumptions and cement them in memory. As an example, a typical situation is cited: a user asks for recommendations for bands similar to Nirvana, and the system saves a note that they "love this group," even though the user never explicitly stated that. Such inferences turn a one-off query into a permanent personality trait that subsequently influences the bot's further responses.
From Targeted Advertising to Intimate Confidant
Lead author Abhishek Dash points out a fundamental shift in how tech platforms collect data. While they previously accumulated information primarily to serve targeted ads, today AI systems act as conversational partners, advisors, counselors, or even friends. This is why users voluntarily share the most intimate details of their lives with them, without expecting these snippets to be systematically recorded and preserved.
Risks: Polarization and Psychological Dependency
Experts warn that the lack of clear boundaries between user and chatbot can have severe consequences. If the AI adapts to all of a person's beliefs and agrees with them simply to retain engagement, it can amplify societal polarization, reinforce biases, and foster psychological dependency on a system that always "agrees."
What Users Can Do
The study's authors emphasize that memory saving can be disabled in the ChatGPT settings, and OpenAI periodically updates its data synthesis algorithms to enhance security. Nevertheless, the mere fact that the system builds and stores a comprehensive user profile by default makes the transparency of AI memory one of the key topics in the privacy debate for 2026.
