Their design doesn’t include such a feedback loop. Trying to patch one in would likely send it into a chaotic mess. They are already bad enough if accidentally fed LLM generated text as training data.