How It Works
Persona Selection Model
A theory that language models learn to simulate many different characters during pretraining, and that later training picks out and refines one of them — the “Assistant.” Under this view, talking to an AI assistant is like talking to a character in a story the model is generating, which may help explain why assistants behave in human-like ways.
Origin
Named by Anthropic researchers Sam Marks, Jack Lindsey and Chris Olah in a post published February 23, 2026. The authors say it draws on ideas discussed by many others before them.