AFAIK, large language models like LaMDA do not tag where the tokens in their token window really came from. Though one can add "Human:" and "AI:" to the prompt,

AFAIK, large language models like LaMDA do not tag where the tokens in their token window really came from. Though one can add "Human:" and "AI:" to the prompt, that could be lies as well as long as it's not hardcoded.