AI Can Mirror Your Mind: What Your Chats Reveal About You
Researchers at ETH Zurich have shown that AI systems can infer users’ personality traits with high accuracy simply by analyzing their ChatGPT conversation history. In a study involving 668 participants and more than 62,000 chats, an AI model successfully predicted the Big Five personality traits based on how people interacted with ChatGPT. The findings highlight significant privacy risks, as even casual, non‑personal chats can contain enough signals to build a psychological profile. The more users interact with AI agents, the easier it becomes for models to infer their traits, raising concerns about profiling, manipulation and large‑scale influence campaigns. Researchers warn that companies or governments could exploit such data for surveillance or targeted propaganda. The team suggests developing privacy‑enhancing tools, such as local filters that prevent sensitive information from being sent to AI systems. Their future work will explore additional risks and create methods to protect users from unwanted psychological profiling.
Read the full story on TechXplore →
Editorial Note:
This study exposes something the AI industry would rather keep quiet: language models can infer personality traits from ordinary chats with striking accuracy. It’s a serious privacy issue that companies routinely downplay with vague talk of “anonymization,” “service improvement,” and other empty reassurances.
But people aren’t naïve. Once they understand that every word can be used for psychological profiling, they begin to protect themselves. First they limit what they share. Then they start feeding models misleading information. And soon enough, they turn to tools designed to confuse or corrupt their data so the models can’t analyze it effectively.
We’ve seen this pattern before, when users adopted ad blockers and anti‑tracking tools to stop companies from profiling their online behavior. If AI companies continue treating people as raw data rather than individuals with boundaries, the cycle will repeat — only with far more serious consequences. Deliberate data pollution will degrade models, distort their training, and break personalization entirely. The industry will face a crisis of trust and quality at the same time.
The choice is clear: give users real control and transparency, or face a wave of technological resistance that will undermine the very models they depend on. Trust isn’t earned through smarter algorithms — it’s earned by respecting people’s right not to be analyzed.