This made something click for me. There really is an issue with humans ingesting our own data. Humans have ended up with many mutually unintelligible languages. If two groups of people train on their own data they gradually start diverging.

But this gets at the heart of the issue - separation. If we ensure that humans and AI are trained on roughly the same data then we will stay connected and be able to understand each other. We may even end up borrowing a few gpt-isms, and that's actually totally fine.