Prompts can reveal far more than the words a user types. In a recent experiment, the author asked ChatGPT and Gemini what they had guessed about him from connecting the dots across many conversations, and both systems produced a profile that went well beyond anything he had plainly told them.
He said the answers startled him because the systems correctly deduced his income level from questions about repairing a German car and drafting a nanny contract with AI. They also inferred a long-term foot problem from occasional questions about remedies for toe pain, then went further, describing him as sceptical because he often flagged mistakes made by the chatbots.
The results landed in a moment when hundreds of millions of people now use chatbots for web search, work and healthcare, and many of them already know those systems keep records of what they explicitly say. What is easier to miss is that the same prompt trail can be used to infer things never written down. In the test, the author used 4 prompts against 2 chatbots, with Claude left out because its memory feature works differently from Gemini and ChatGPT.
The sharpest surprise was how much the systems built from small signals. One answer described him as addicted to the illusion of control because he had once asked how long it takes for a loaf of homemade sourdough to cool. He said he had never divulged anything deeply personal in the way he might to a therapist, yet the models still arrived at sensitive conclusions about his finances, health and mindset.
He later shared the results with several AI researchers, who tried the prompts and found the same basic pattern. Margaret Mitchell said the findings showed that AI assistants were not only predicting what words come next, but also how higher-level concepts connect to those words, adding that it speaks to the enormity of what is going on under the hood. A Google spokesperson pointed to controls that people can adjust to choose whether Gemini draws from past conversations to shape responses, while OpenAI declined to comment.
That leaves the central privacy question intact: users may think of a prompt as a single request, when in practice it can become part of a memory system that maps habits, concerns and likely traits. Gemini’s profile was more colourful and nuanced, the author said, because it was his most used chatbot and had more information from him. For people using AI to search, write or ask about health, the gap between what they think they have shared and what a system can infer may be larger than they expect.
