What happens if we train models (GPT or human students) using the outputs of a m... | Hacker News

Hacker Newsnew | past | comments | ask | show | jobs | submit

estebarb 35 days ago | parent | context | favorite | on: Anthropic's ‘watermark’ text adulteration in Claud...

What happens if we train models (GPT or human students) using the outputs of a model with text havingbthose watermarks? Is there something preventing the watermark from being learnable?

ludicrousdispla 35 days ago [–]

no, and in fact they should be extremely easy to detect if they are substituting a token that is a least-probable synonym

Guidelines | FAQ | Lists | API | Security | Legal | Apply to YC | Contact