In early 2026, researchers at ETH Zurich published a study that should have been bigger news than it was. They analyzed 62,000 ChatGPT conversations from 668 users, then tried to predict each user's Big Five personality traits from those conversations alone — without asking them any questions about themselves.
It worked. Not perfectly, but significantly better than chance across multiple dimensions, with extraversion and openness showing the strongest signal.
The finding confirms something that researchers in computational personality science have been accumulating evidence for over the past decade: the way you communicate, the topics you engage with, and the texture of how you reason carry consistent personality information — information you're not deliberately sharing and can't easily manage.
What the ETH Zurich study actually found
The study gave users a standard Big Five questionnaire and collected their ChatGPT conversation histories with consent. Then researchers trained models to predict Big Five scores from conversation data alone — text features, topic patterns, linguistic style, and interaction sequences.
The results were uneven across traits, in a way that makes sense:
Extraversion was the most predictable trait. Extraverts engage more frequently, write longer messages, and show a different conversational rhythm than introverts — more back-and-forth, more socially-framed requests. The signal is strong enough that conversation features significantly outperformed random chance.
Openness was also strongly predictive. Topic breadth is the primary signal — the range of domains a user engages with across their conversation history is a remarkably stable indicator of their openness to experience score. Curious people are curious about many things, and it shows in what they bring to an AI.
Conscientiousness showed moderate predictability. Indicators included prompt specificity, follow-through on multi-step tasks within conversations, and the degree to which users structured their requests clearly before asking. These signals are real but noisier than extraversion or openness.
Agreeableness was weakly but consistently predictable. Social register in prompting style — how warmly requests are framed, whether users include context about impact on others — carries a signal, but it's subtle.
Neuroticism was the hardest to predict reliably. Emotional volatility and negative affect don't leave a clean signature in AI conversations the way they might in social media data — partly because people are often using AI specifically to manage difficult situations, which creates confounding patterns.
The researchers were careful about the limits of their findings. This is not surveillance. A single conversation reveals almost nothing. The signal emerges from aggregate patterns across hundreds of interactions, and even then it's probabilistic, not deterministic.
Why this data source is different from a personality test
The standard objection to personality self-report is well-documented: you answer based on how you see yourself and how you want to be seen, not necessarily how you actually behave. Research on informant-rated personality — having people who know you well rate your traits — consistently shows that observers predict outcomes better than self-raters, because they watch behavior rather than intentions.
AI conversation history is not the same as observer ratings, but it shares a key property: it's generated when you're not trying to describe yourself. You were asking for help with something, thinking through a problem, or working on a project. The way you went about it wasn't crafted for an audience evaluating your personality. That's what makes it a different category of evidence.
There's a technical name for this in psychometrics: behavioral validity. Self-report has poor behavioral validity for many traits because self-concept and behavior diverge, especially for traits where there's a socially desirable answer. Behavioral data bypasses this problem by measuring what you actually did rather than how you rate what you do.
Which signals leak out — and why
The patterns researchers find in AI conversation data map predictably onto what we know about each trait:
Extraversion shows in social framing. Extraverts are more likely to write about other people, frame problems in interpersonal terms, and use AI as a thinking partner for social situations. Their conversation histories have more characters in them — more names, more relationship dynamics, more "how should I respond to..." queries.
Openness shows in topic range and intellectual style. High-openness users range across philosophy, science, art, technology, history, and creative work — often within the same week. They're more likely to ask exploratory questions ("what are the arguments for and against...") than convergent ones ("give me the answer to..."). The conversation history of a high-openness user looks qualitatively different from that of a low-openness user even before any statistical analysis.
Conscientiousness shows in how people structure requests and manage task completion. High-conscientiousness users tend to provide extensive context upfront, break complex asks into clearly specified steps, and follow up on their own prior conversations. They're more likely to return to a topic to close it out rather than leaving it open-ended.
Agreeableness shows in interpersonal framing and social register. Agreeable users tend to frame requests in terms of impact on others, maintain warm phrasing even with a machine, and ask more questions about how to navigate interpersonal situations graciously rather than effectively.
Honesty-Humility (from the HEXACO framework, not measured in the ETH Zurich study but relevant here) shows in how people reason through ethical tradeoffs. When presented with situations involving competing interests, high-H-H users consistently engage with the fairness dimension; low-H-H users more readily explore strategic framings that sideline it.
Attachment style shows in what proportion of AI use is dedicated to processing relational uncertainty — and in how those conversations are framed. Anxious attachment creates a characteristic pattern of seeking interpretation of others' behavior, returning repeatedly to ambiguous relational situations, and asking questions that reveal genuine ambivalence about closeness.
What this doesn't mean
The ETH Zurich study — and the broader research tradition it sits in — is sometimes misread in privacy-threatening directions. A few clarifications are worth making explicit.
It doesn't mean OpenAI can read your personality. Personality inference from AI conversations requires significant data, appropriate models, and deliberate analysis. It's not happening automatically in the background, and standard AI interactions are not producing personality profiles without your knowledge.
It doesn't mean the predictions are deterministic. The accuracy gains over chance are real but modest. A single conversation tells you almost nothing. Long histories produce better signal, but the predictions are still probabilistic — they describe a person who tends toward a trait, not a person who has a fixed quantity of it.
It doesn't mean you should change how you use AI. If you find yourself wanting to curate your AI use to manage what your history might reveal — you've mostly just added another layer of self-presentation to your data, which researchers could probably detect.
The more useful response is curiosity: if your AI conversations carry personality signal, and if you've been generating those conversations for a year or more, what does your specific history actually show?
From research finding to personal insight
The ETH Zurich study established the principle. Memrov applies it as a consumer product.
You export your conversation history from ChatGPT, Claude, or Gemini. Memrov reads the behavioral patterns across your full history and generates a personality profile across six validated frameworks: Big Five, HEXACO, attachment style, Schwartz values, Dark Triad, and motivation patterns.
The output isn't a statistical model output — it's a narrative reading that interprets what those patterns mean for how you actually operate. It draws on the same behavioral signals the researchers identified, plus additional frameworks that extend well beyond what a single academic study covers.
Your raw export is stored for seven days maximum, then permanently deleted. The derived reading stays.
The data is already there. The conversations have already happened. The question is just whether you want to see what they've been accumulating.
Upload your AI conversation history and get your free personality reading →