GPT-4 is, but ChatGPT is fine-tuned to emit sentences that get rated well (by humans, and by raters trained to mimic human evaluation) in a conversational agent context.