Twelve Prompts to Know You: The Inversion Point Where AI Models Users Better Than They Model Themselves

Every conversation with a large language model is bidirectional: the user extracts answers, and the system extracts behavioral signal. We introduce the inversion point, the threshold at which an AI system predicts a user's behavior more accurately than the user's own self-model can. Analyzing 128,575 messages from 7,159 users in the OpenAssistant OASST2 corpus, we find this threshold is 12 messages. Specifically: (1) stable behavioral personas emerge after 10 messages (cosine similarity 0.91); (2) bigram topic-transition models predict the next topic at 45.4% accuracy, 7.7x above chance; (3) a single message identifies its author from 100 candidates at 31.5x above random; (4) when confident, the system is correct 97.6% of the time; and (5) users grow more predictable with continued use. These results use only statistical methods, making them a conservative floor. We propose governance mechanisms including the right to unpredictability and mandatory diversity injection beyond the inversion point.