Home | Notifications | New Note | Local | Federated | Search | Logout

Note Detail

princess pancake :butterfly_:​:neofox_lesbian:@natty@astolfo.social (2026-08-18 05:10:33)
LLMs watermarking text using stylometry has some fucking horrifying implications, and it gets worse the more you think about it (from the least concerning to the worst): 
- since it is a statistical analysis trick, real threat actors will trivially bypass this with a Python script
- it's trivial to forge and frame someone by reverse engineering the pattern and framing someone 
- you don't know which model generated the text so you gotta check against a wide spectrum providers, which sounds like a lucrative business for those that provide it
- conveniently, AI companies can use it to scrape only human written content instead of inbreeding 
- it is purely probabilistic and good luck explaining that to cops or courts that a 200 word text has an 80% true positive rate