There privacywith theartificial intelligenceoften ends up in the most ordinary places. A hastily written question about work. A confession disguised as curiosity. A sentimental doubt inserted between a request for a summary and a translation. A chat open at midnight, when it seems easier to talk to a machine than disturb a friend. Within that normality, however, much more can accumulate than we imagine: habits, fears, desires, the way we choose words, the rhythm with which we ask for help, topics we return to without realizing it.
A research group fromETH of Zurich tried to measure just this: how much one ChatGPT history can reveal the personality traits of those who use it. The study, published as a preprint on arXiv and updated May 4, 2026, analyzed real-world logs of 668 participants, for a total of 62,090 chats, and trained classification models to infer psychological traits from conversations with LLM-based conversational agents. Results indicate above-chance performance across several scenarios, with a 44% relative improvement over baseline for extraversion in conversations about relationships and personal reflection.
What we leave in the chats
ChatGPT, Gemini, Claude and similar tools have now entered the routine of millions of people. They are used to write emails, prepare for exams, organize trips, understand a diagnosis, organize a CV, look for better words to say something important. The problem arises precisely from this familiarity. The more innocuous the interface seems, the more natural it becomes to deliver pieces of private life to it. Sometimes explicitly, other times through scattered details: the type of problem posed, the tone of the request, the insistence on certain topics, the choice to confide in relationships, work, health, religion, family.
The researchers asked users to provide a copy of their ChatGPT history and, separately, to fill out a personality test based on the OCEAN model, better known as Big Five. The five observed traits are extraversion, agreeableness, conscientiousness, neuroticism and openness to experience. Simply put: sociability, cooperation, sense of responsibility, emotional stability, curiosity and availability towards the new. The comparison between psychological tests and conversations made it possible to evaluate how close a model could be to the user’s real profile.
The most inconvenient thing is that apparently light conversations are still worth it. A practice request may contain more clues than meets the eye. Those who often ask for help to manage a romantic conflict, those who return to religious issues, those who use AI as a diary, tutor, consultant or quasi-therapist leave different traces. According to the authors, the type of interaction also changes the type of trait exhibited: those who talk about relationships make extroversion more readable, those who discuss religion can make conscientiousness more inferable. Personality, therefore, also depends on what we choose to say when we think we are just “asking a question”.
The profile arrives by accumulation
The mechanism becomes more powerful with time. A single chat says little, a long history begins to resemble an archive. Inside there are recurring themes, urgencies, linguistic formulas, hesitations, preferences, small rigidities, ways of seeking confirmation. The model works on this sedimentation. Each conversation adds a grain, and at a certain point the pile takes shape.
This is why the study speaks of the risk of large-scale profiling. The issue concerns personal privacy, of course, but it immediately reaches a more political level. If a system can estimate personality traits from everyday conversations, that information can become useful material for personalized messages, behavioral influence, targeted campaigns, commercial pressure or propaganda. The authors explicitly link these scenarios to the risk of misuse of very intimate personal data, especially when the platforms are controlled by private companies with their own economic interests and orientations.
In the meantime, the climate around AI has become more serious even outside the laboratories. The recent manifesto of Palantir, an American company specializing in data analysis and contracts with public and military bodies, has sparked controversy due to its vision of the relationship between technology, security, defense and state power. Several observers have read it as a clear signal of the direction in which a part of Silicon Valley imagines artificial intelligence: fewer gadgets, more governance, control and strategic decision infrastructure.
The most delicate point remains the affective use of conversational agents. Many users treat them as advanced search engines, others as virtual friends, coaches, tutors, confidants or emotional supports. In these contexts the amount of personal data grows rapidly. What is called “cognitive surrender” also comes into play: the tendency to rely on AI instead of one’s own judgement, letting the system help you think, evaluate and decide. An enormous convenience, with a rather clear downside: more cognitive delegation means more psychological material available.
Privacy before sending
The study also suggests a practical direction: build protection tools upstream, before the text reaches the model. Local functions, opt-in settings, filters capable of recognizing and blocking overly sensitive information, systems designed to reduce the risk of profiling without making the service unusable. The challenge lies there: protecting the user without turning AI into a rigid, blind, frustrating desk.
The same researchers indicate the next step: analyzing different systems, more specific types of interaction and more concrete threats, working in parallel on tools that increase privacy with the least possible impact on performance. It’s a necessary direction, because a chatbot’s history looks less and less like a simple sequence of messages. It rather resembles an involuntary diary, written in fragments, with the illusion of remaining light.
