"The human evaluation results indicate that the response generated from DialoGPT is comparable to human response quality under a single-turn conversation Turing test."

Paper here: https://arxiv.org/abs/1911.00536