Invisible AI watermarks sound like the kind of thing dreamed up in an exceptionally boring episode of Black Mirror.
Anthropic is introducing invisible watermarks across supported Claude models, alongside provenance metadata for generated files, partly in response to transparency requirements under the EU AI Act.
So what is this watermark, and does it even matter?
For text, the mark is essentially an invisible statistical signal woven into Claude’s output. It doesn’t appear on the page, shouldn’t noticeably change the writing, and can survive things like copying, pasting and light edits.
The idea is simple enough: give people a way of identifying text that has passed through Claude.
Where it gets interesting is that “passed through Claude” doesn’t necessarily mean “written by Claude.”
You could write an entire article yourself, put it into Claude for a proofread or a bit of editing, and the returned version may still carry the watermark.
That means detecting one doesn’t automatically prove AI wrote the ideas, research or original copy. It may simply prove Claude had a hand in the final version.
For anyone using AI as a very expensive spellchecker, that distinction is worth knowing.
The watermark isn’t bulletproof either. Heavy rewriting, translation or substantial editing can weaken it, while very short pieces of text may not contain enough information to reliably detect anything.
Claude-generated images and supported files use a different system, with C2PA provenance metadata recording that Claude was involved in creating or modifying them. That metadata can also disappear through screenshots or file conversions.
So, does any of this matter?
Yes, mostly because of what happens after something gets detected.
A watermark can tell you Claude was involved somewhere in the process. It can’t necessarily tell you how much.
And as someone who spends a fair amount of time writing, editing, obsessing over Oxford commas and occasionally staring at one bad sentence for far longer than is reasonable, this is the part that got me: you don’t actually have to ask Claude to write something for it to leave its fingerprints on your work.
And if someone points to a Claude watermark and confidently declares that an entire article, essay or piece of copy was “written by AI”, it is worth asking what Claude actually did.
Because the answer might be considerably less exciting.
Someone may have just wanted a little help with an Oxford comma.
