claude's text is getting an invisible watermark
anthropic is adding a watermark to claude's text to comply with eu rules. it's a pattern in word choice, not hidden characters, and readers can't spot it.
what's happening
anthropic says future claude models will generate text that carries a watermark, to comply with eu rules.
it's not hidden characters. the model's word choices follow a pattern that someone with the key can check. to a reader, the text looks the same.
anthropic says the watermark carries no identifying information and can't be traced to a person, organisation or chat.
the rollout
per the decoder, anthropic emailed api customers that claude mythos 5, opus 4.7, opus 4.6 and sonnet 4.6 get the watermark, with newer models already carrying it.
opus 4.6 was later pushed to 16 november, alongside opus 4.5, sonnet 4.5 and haiku 4.5. on some older models the mark only applies when sampling settings are left at their defaults.
not only claude
anthropic says other major model developers signed the same eu code of practice and will add their own watermarks.
who can check it
not fans, for now. checking requires the key.
the policy part
anthropic's new usage policy has a section against deceptive campaigns and artificial activity, including amplifying content through fake accounts or posts.
why it matters for a creator business
if your dms run on one of these models, the text going out will be marked. today nobody on the fan side can check it. that can change.
the daily post, by email. nothing else.
one email a day max. unsubscribe anytime.