Claude Is About to Watermark Your Drafts. Here’s What That Actually Means.

You will not see Claude’s new watermark. There is no hidden character, no extra token, no footer that says “written by AI.” Future Claude models will leave a statistical pattern in the low-stakes word choices. Overcast or grey. The sentence still means the same thing. A detector with Anthropic’s key can later say Claude was likely involved.

That is the whole trick. And it is not live on the Claude you are chatting with today. Anthropic published the plan on August 14. The mark arrives with future models, then older ones over the coming months.

Why this is happening

The EU AI Act’s marking obligation kicked in on August 2, 2026. In July, Anthropic signed the EU Code of Practice on Transparency of AI-Generated Content, along with other major providers and about 190 signatories total. Everyone who signed is supposed to mark AI-generated text. They will each do it their own way.

Anthropic is applying the watermark worldwide at launch. Not because the Act covers Kansas. Because they do not yet have a durable way to scope it by region. They say they will keep evaluating that.

Source: Anthropic’s watermark explainer.

How it actually works

Language models pick the next word from a list of candidates. When two options are equally fine, the choice is usually random. Watermarking changes the source of that randomness. Instead of an arbitrary random number, the model uses a secret key plus a few preceding words to settle the pick. Over a long enough passage, those low-stakes choices form a pattern. Readers cannot see it. Anyone with the key can check whether the sequence looks like Claude using that key.

The method is a version of Google DeepMind’s SynthID-Text. Nothing is inserted into the text. No extra tokens, so it does not cost more. Anthropic says internal testing and a human rater study found no practical impact on content, creativity, or readability. The original SynthID-Text paper reported the same kind of result on Gemini traffic: no statistically significant change in thumbs-up or thumbs-down.

It stays sparse where the next word is not optional. Factual lines (Newton’s book is Principia Mathematica, not a synonym). Code that would break if you swapped a method name. Comments can carry a mark. The actual code usually cannot.

What a hit can and cannot prove

A watermark can only say Claude was likely involved. Wrote it, or heavily edited it. It cannot tell those two apart. It cannot detect another lab’s watermark. It cannot confirm the text is human. It carries no user, org, or chat identity. There is no personal data in the key.

Short samples are weak. More word choices, more confidence. Light proofreading of your own draft may leave too few Claude-chosen words to register. A full rewrite, every word replaced, removes it. Anthropic’s own line: at that point it is arguable whether the text is still AI-generated.

Translations go the other way. If Claude translates the piece, every word is Claude’s, so the watermark is there.

Files are a different system. When Claude produces a supported PNG, JPG, or SVG, it attaches C2PA signed provenance metadata. That is a note in the file, not a pattern in the pixels. Strip the metadata and the label is gone. Anthropic says it will offer a drop-a-file checker for that. Details later.

The help article is the cleaner version of the same limits: How Claude marks AI-generated content.

Gemini’s contrast: a sparkle you can hide

Same week, Google shipped a toggle. Gemini 3.7 Flash is now in the regular chat picker, not just Spark. A new “Media watermark” setting can hide the visible corner sparkle on generated images, video, and songs. Default is on. Invisible SynthID and C2PA stay either way. The toggle is not available in countries that require a visible mark.

That is the pairing. Gemini will let you hide a logo in the corner. Claude’s text mark lives in the words. You cannot flip it off.

Source: The Verge on Gemini’s visible-watermark toggle.

Weekend checklist

• If you use a future Claude model to polish or translate, assume a detector can later say “Claude was involved,” not “Claude wrote this.”
• A light grammar pass is less likely to leave a usable mark than a rewrite or a full translation.
• Code and tight factual copy will carry a thinner signal. Do not treat a miss as proof a human typed it.
• You cannot check a file today. The detection API is “soon.” Details are not set.
• PNG, JPG, and SVG from Claude will get C2PA metadata, not a text-style watermark in the image.
• Older models (launched before August 2, 2026) are on a transition clock. Watermarking for those rolls out over the coming months.

What we still do not know

When the detection API ships. Who can call it. What a score looks like. How fast older models get the mark. Whether Anthropic ever finds a way to keep it EU-only. Whether other signatories’ watermarks will be interoperable (today’s answer: no, different keys, maybe different methods).

Until then, the practical rule is simple. Future Claude text can carry a silent “we were here.” It will not name you. It will not prove authorship. And if you asked Claude to translate the whole thing, that silent mark is part of the draft.