Depends on how you define privacy. If definitely accessing the data to train it was illegal/wrong, then you can't. But if that is ok, then the question is to wh

Depends on how you define privacy. If definitely accessing the data to train it was illegal/wrong, then you can't. But if that is ok, then the question is to what application will the LLM be applied? A generalized chat interface probably would be able to regurgitate private information. But things like reasoning engine, or recommendation system plugged into other downstream systems where the raw output isn't expose very much could work. There are a hundred new startups everyday it seems that are some kind of wrapper on top of an LLM applied to a niche use case. Many of these should be able to preserve privacy.