[tgv]

But when two sets of fingerprints, are different, you can be fairly sure they're from different people. But when the percentage of some features is 20% in one text, and 30% in another, you still can't conclude anything. I write in different registers in contexts such as personal emails, professional emails to a large group, professional emails to a direct colleague, a quick post on the internet, an 'app' to a friend in another country, a text message on a phone, etc. I even write them in different languages. It's hard to imagine there's a well-defined, properly grounded model that can unite those yet distinguish them from written output by other people.

And now LLMs are going to add more noise to these features...