Also, while you can generate an interview that might sound like the individual in question has given it, insofar as an interview is there to obtain new information, an LLM trained on their other speech isn’t going to be very useful. The LLM can’t magic new information about the guy out of the air.

Maybe it’d be possible to combine information from one person’s speech and another’s to obtain an insight that other humans didn’t pick up on. Like, okay, say that you had interviews with Mormons, and you have the speech of someone who has never publicly announced that they are Mormon but was Mormon and had speech that has common characteristics with other Mormons. Maybe the LLM could, based on those commonalities, generate a response where the LLM says that the individual in question is Mormon. That might have some value in that other humans might not be able to pick up on the commonalities.

But I’m really stretching to find some kind of way in which doing interviews is a reasonable application.