A lot of the replies seem to read this as a writing style thing, but from what I can tell the paper is mostly about content. The model pulls out what people say about themselves (city, job, hobbies, trips, events) and matches on that. So rewording your posts through an LLM probably doesn't help much, the facts are still there.
The pseudoanonymous reputation carrying we so like here in this community become an issue. Carrying one name across X, Nostr, forums and here is how people build trust, but it also looks like what makes the matching easy, according to paper.
Single details are innocent, but the combination... I think it's something like 33 bits to narrow it down to one person? Not sure what the practical fix is besides just saying less.
A lot of the replies seem to read this as a writing style thing, but from what I can tell the paper is mostly about content. The model pulls out what people say about themselves (city, job, hobbies, trips, events) and matches on that. So rewording your posts through an LLM probably doesn't help much, the facts are still there.
The pseudoanonymous reputation carrying we so like here in this community become an issue. Carrying one name across X, Nostr, forums and here is how people build trust, but it also looks like what makes the matching easy, according to paper.
Single details are innocent, but the combination... I think it's something like 33 bits to narrow it down to one person? Not sure what the practical fix is besides just saying less.