DeepL for example already used machine learning for their translations. Over the years they accumulated a massive amount of high quality data thanks to language experts they employ.
Their current model is much better, because their training data is much better than feeding everything in.