“Major AI labs like Anthropic, Google (for Bard), and OpenAI almost certainly use a post-generation classifier to filter for safety, in addition to safety measures within the RLHF pipeline.”