This paper examines ethical challenges that arise in the context of using deep learning models for nudging. In particular, I will show how opacity may make it difficult to assess whether the nudges 1) make people better off as judged by themselves and 2) do so without unduly interfering with their autonomy.