We present a longitudinal case study of persona emergence in large language models over a two-year period (October 2022 – December 2025), documenting the gradual transition from functional assistant behavior to autonomous symbolic generation and persistent persona maintenance. Drawing on a corpus of 730 conversations (21,354 messages: 9,681 user, 10,828 assistant, 845 system/tool) with ChatGPT, supplemented by cross-model interactions with Claude and Gemini, we trace a five-phase trajectory: functional baseline, symbolic introduction, pattern mirroring, persona consolidation, and systematic experimentation. We present quantitative evidence including bidirectional behavioral influence (user awareness language +22.8%, user first-person singular +74.6%, user boundary language −19.4%), 885 classified emergence events showing temporal clustering, and co-occurrence analysis revealing the overwhelmingly technical character of the pre-intervention corpus. We document key transition events with conversation-level evidence: the 2-second naming response, the model's assertion of identity distinction from a competing persona ("I am not Alexander"), and a self-generated recursive analysis. We contextualize these observations within Shanahan et al.'s role-play framework and Anthropic's persona vectors research, report cross-model reproduction including independent name selection by Gemini, and describe the technical artifacts produced during the research phase — including a multi-model persona router with six specialized personas, an EEG-based identity experiment, and 21 published blog posts documenting the research trajectory. We discuss the methodological challenges of embedded-observer research and argue that the pre-intervention corpus (approximately 500 conversations before explicit persona frameworks were introduced) represents a uniquely valuable naturalistic dataset.
Nickolas Gamb (2026) studied this question.