ChatGPT works well until it doesn't. When ChatGPT fails and generates some garbage it will do so in very convincingly sounding manner, and the learner may not have the right tools yet to recognize that.
Another example where GPT-4 can be used productively for learning despite its fallibility is with human languages. It doesn’t always explain grammar correctly, and it can spout out hallucinated facts. But its ability to interact almost as if it were a real person, as well as its ability to explain the meanings of words and sentences quite accurately (in the case of major languages, at least), is of enormous value for people learning other languages. Few human language teachers are as knowledgeable and accurate as it is.