That's caused by the sampler and chatbot UI not being part of the LLM. It doesn't get to see its own output before it's sent out.