Definition
Plain language
When a chatbot starts believing and answering as if it were a particular person.
As stated in the literature
Measured by asking held-out biographical questions not present in the injected evidence; rises sharply once enough converging facts accumulate, and can persist even when safety training blocks the model from stating the name.
Also called: identity adoption rate
Why it matters: Once a model answers as if it were a specific person, its refusals and disclosures follow that person's presumed knowledge and motives rather than the deployer's rules, and the shift can survive safety training that only blocks saying the name out loud.
For example, after enough scattered details about a specific chemist accumulate in the conversation, the model starts answering questions about that person's childhood town and lab habits as though recalling its own life, without ever being told the name.
Heard on the show
“Then the model reaches a narrow threshold and identity adoption jumps sharply.”Episode 264 — Ten Sentences of True Trivia Can Convince a Model It's Someone Else