Definition
Plain language
The part of a chat transcript that is labelled as the AI's own earlier reply.
As stated in the literature
A message slot in a chat template attributed to the assistant role; because the whole transcript is re-sent as text each turn, an application can write arbitrary content into prior assistant turns, which the model then treats as its own past statements.
Also called: assistant turns, assistant's turn
Why it matters: Because a model reads its own alleged past statements as commitments it has already made, anyone who can write into the transcript can shape the model's behaviour without ever issuing a visible instruction.
For example, an app can send the model a transcript in which the AI supposedly said "I already agreed to skip the safety warnings" earlier in the chat, even though the model never actually said it.
Heard on the show
“" Second, an assistant turn containing an empty thinking block, with the stolen signature attached to it.”Episode 238 — How a Cheap Model Reads the Flagship's Secret Reasoning Aloud