Claude gets invisible watermarks

Future Claude models will watermark their text to show AI involvement. This complies with the EU AI Act without affecting quality or cost.

No visible changes to text

The watermark is invisible to readers. Nothing is added, there are no hidden characters, and it costs no extra tokens.

“The difference between watermarked and un-watermarked text will not be distinguishable to readers”

How language models choose words

Claude generates one word at a time, choosing from likely candidates. After "cold and," it might pick "overcast" or "grey," not "sugary."

Watermarks use low-stakes choices

When multiple words work equally well, watermarking creates a pattern. The pattern is invisible but detectable with a key.

A different source of randomness

Instead of arbitrary random numbers, the watermark uses a key and previous words to pick the next word. Choices remain random but follow a pattern.

No impact on quality

Internal testing shows no effect on content, creativity, or readability. Google DeepMind found no difference in user ratings between watermarked and unwatermarked text.

The Monopoly analogy

It's like using digits of pi instead of dice rolls in Monopoly. The game plays the same, but you can verify the pattern afterward.

Based on SynthID-Text

Claude uses a version of Google DeepMind's SynthID-Text approach, published in Nature in 2024. It builds on Scott Aaronson's 2022 proposal.

Watermarking has limits

It only shows likelihood Claude was involved, not whether text is human-written. It doesn't work well on short passages with fewer word choices.

Factual text has sparse watermarks

When accuracy matters, there are fewer choices. "Isaac Newton's most famous work was called Principia Mathematica" has only one right answer.

Proofreading barely shows watermarks

When Claude only edits grammar and punctuation, nearly all words remain human-written. There's little for the watermark to attach to.

Code gets minimal watermarking

Code must be exact, so watermarks rarely apply. They only appear in arbitrary choices like comments, not in functional code.

No slowdown or extra cost

Watermarking has negligible speed impact and produces no extra tokens. The model costs the same to serve and use.

Watermarks don't identify users

The watermark applies only to Claude's outputs. It contains no information about individual users, organizations, or specific chats.

Required by EU law

Anthropic signed the EU Code of Practice in July 2026 with 190 signatories. It's applied globally because regional scoping isn't ready yet.

Detection API in private preview

A detection API is available to regulators, law enforcement, media, and enterprises for compliance. Access will expand over time.

Images get C2PA credentials

Claude adds cryptographically signed metadata to image files using the C2PA standard. It's readable by any C2PA-aware tool.

Editing can remove watermarks

Light editing probably won't remove the watermark completely. A complete rewrite where every word is replaced will.

Different from AI detectors

Tools like Pangram look for AI phrasing patterns without the key. They spot tells like "this isn't X, it's Y" or overuse of "quietly."

Ownership rights unchanged

Watermarks only test if Claude processed the content. They don't affect ownership, authorship, or user rights under Anthropic's terms.