The Smart Compose paper does say that it was trained on user emails, and that each user gets a model trained for their personal style.