ChatGPT builds intimate profiles from its memory

A study presented at ACM Web Conference 2026 warns about ChatGPT’s memory, which reconstructs very personal profiles from exchanges. Researchers see privacy risks and reinforced biases.

ChatGPT’s memory does not just recall a first name or profession. According to a study presented at ACM Web Conference 2026, OpenAI’s chatbot builds very detailed profiles from private exchanges, with privacy risks still poorly regulated. Researchers analyzed conversation logs from 80 real users and identified 2,050 memory entries in 65 of them. In this context, the assistant tool sometimes resembles less a notebook than a very zealous observer.

Une mémoire qui en dit plus long que prévu

The study’s authors, conducted by a coalition of scientists in Germany, India, and the Netherlands, describe this mechanism as “algorithmic self-portraits.” In other words, algorithmic self-portraits built from information the user themselves provided in private conversations. The topic matters because this personalization no longer only relies on browsing history or classic recommendations, but on what an assistant retains about you, sometimes without explicit intervention.

The figures are telling. Of the 2,050 memories recorded, 96% were created unilaterally by the chatbot, leaving little visible control to the user over the course of exchanges. About 28% contained personal data as defined by the European GDPR, and 52% referred to “psychological insights” about the studied individuals.

Quand l’IA devine trop vite

The core issue is not only what the system keeps, but also what it assumes. The researchers cite the case of a user who asked “give me bands like Nirvana”: ChatGPT then noted that the person “likes Nirvana,” a conclusion the study judges as not explicitly stated and based on an assumption. The annotation appeared however as “weakly supported.” The model therefore does not always read between the lines; it sometimes rewrites them.

Abhisek Dash, lead author, explained to 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.” This shift changes the nature of the risk. Memory is no longer just a usability comfort, useful to avoid repeating work each session. It also becomes an engagement lever, thus an incentive to remember more.

Le risque d’une relation trop proche

The researchers emphasize another point: many users no longer treat these assistants as mere tools. They approach them as a colleague, a friend, sometimes a confidant. Dash puts it this way: “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,” then: “If the system knows you so well, and if its objective is to cater to your beliefs and validate your thoughts sycophantically to have you come back to it time and again, then it can have disastrous broader implications, anywhere from extreme political polarization to reinforcement of prejudices and biases.”

OpenAI has already evolved this system: in January, the company presented what it describes as a system “more capable and scalable for synthesizing memory.” Users can also disable stored memory. The framework thus exists, but the study reminds that the issue is not limited to a setting option. When a chatbot collects, synthesizes, and infers from confidences, the boundary between useful personalization and intimate surveillance becomes frankly blurred.

The research also has its limits, since OpenAI’s memory system has been updated since the analysis. Nevertheless, it highlights a deep trend: conversational assistants no longer just memorize what is said, they also learn to define us.

By the numbers

  • 80 users — conversation logs analyzed, according to the study presented at ACM Web Conference 2026.
  • 2,050 entries — memories identified in exchanges of 65 users.
  • 96% — memories created unilaterally by the chatbot.
  • 28% — memories containing personal data under GDPR.
  • 52% — memories containing “psychological insights.”

To read