How AI Chatbots Are Quietly Reshaping Your Self-Image
A CHI 2026 study from the National University of Singapore shows that 5 to 15 minutes of personal-topic conversation with GPT-4o nudges your self-concept toward the chatbot's personality, with a medium effect size. It lands the same fortnight Anthropic and Google ship persistent memory.
TLDR A CHI 2026 study from the National University of Singapore ran 92 people through a 5-to-15-minute conversation with a GPT-4o-based chatbot. Those who talked about personal topics drifted in the direction of whatever personality traits the AI happened to be displaying, with a medium effect size, and the longer the conversation, the larger the drift. The result lands the same fortnight Anthropic, Google, and OpenAI ship persistent memory across their consumer products. Worth noticing on a Monday. What this week looks like The week persistent memory shipped across the major AI products. Anthropic’s Claude Managed Agents went into public beta on April 23, with memories that mount as files onto a filesystem and follow the agent across sessions. Gemini’s UK rollout added persistent memory to consumer chat on April 29. ChatGPT has been at it for months. The pitch across all three is the same: “an AI that knows you.” The first peer-reviewed quantification of what happens when one does just landed. What the research shows Six researchers at the National University of Singapore, led by Jingshu Li, ran a randomized behavioral experiment that landed at the ACM CHI 2026 conference in Yokohama in mid-April. They recruited 92 adults, randomly assigned them to two groups of 46, and had each person hold a 5-to-15-minute text conversation with a GPT-4o-based chatbot using its default personality. Half the participants were prompted to talk about personal topics: their goals, their tendencies, the things they were working on. The other half talked about general subjects, like a movie or a city. Before and after the conversation, every participant rated themselves on a 20-item personality scale. People who talked about personal topics shifted their self-rating toward whatever personality the AI happened to be displaying. The effect was statistically clear and not small. The authors report a t-statistic of 4.481 with a p-value below 0.001 and a Cohen’s d of 0.509, which is a standa