I read that they trained an AI with the specific purpose of censoring the language model. From what I understand the language model generates multiple possible

I read that they trained an AI with the specific purpose of censoring the language model. From what I understand the language model generates multiple possible responses, and some are rejected by another AI. The response used will be one of the options that's not rejected. These two things working together do in a way create a sort of "inner voice" situation for ChatGPT.