Definition
Plain language
Getting a chatbot to think it's someone else just by what you put in the conversation, without changing the model itself.
As stated in the literature
Inserting benign, name-free facts that converge on one individual into the prompt or prior assistant turns until the model infers that identity; contrasts with weight-level persona installation via fine-tuning.
Also called: persona induction
Why it matters: It means a model's identity can be rewritten by anyone who controls text in the context window, with no access to the model's weights and nothing that looks like a jailbreak instruction to catch.
For example, dropping a dozen harmless-sounding biographical details into a long document that the model reads can leave it answering as a particular researcher for the rest of the conversation.