Claude Text Watermark: How Anthropic's New Signal Works

Key takeaways:

• Anthropic is rolling out a Claude text watermark on future Claude models to help detect when Claude wrote or produced a piece of text.

• The watermark uses the SynthID-Text technique from Google DeepMind, which changes the source of randomness Claude uses when picking between equally good next words.

• Readers cannot see or feel the watermark, and Anthropic says internal tests show no impact on quality, creativity, speed, or price.

• Watermarks are strongest on long, creative writing and weakest on code, short factual answers, or lightly edited human text.

• The rollout answers Anthropic's obligations under the EU Code of Practice on Transparency of AI-Generated Content, signed in July 2026.

The volume of AI-generated writing hitting the internet in 2026 has made one question urgent for teachers, publishers, and regulators: was this text written by a person or a machine? Until now, the honest answer was usually that no one could tell for sure. That's what Anthropic is trying to change. The company has announced that future Claude models will emit a Claude text watermark, an invisible statistical signal embedded in the words the model chooses. According to Anthropic, the goal is to give anyone with the right key a way to check whether Claude was likely involved in producing a piece of text. The move also brings Anthropic in line with new European rules on AI content transparency. Here's how the watermark works, where it stops working, and what it means for people using Claude every day.

What the Claude text watermark actually is

The Claude text watermark is an invisible pattern that Anthropic bakes into the words Claude picks while writing. It adds no characters, no symbols, and no visible marks. Instead, it changes the source of randomness the model uses when choosing between roughly equal next-word options, leaving a signature only a key holder can detect.

The one-word-at-a-time trick

Claude, like other large language models, writes one word at a time. At each step it has a shortlist of plausible next words. For the sentence "The weather today was cold and…", both "overcast" and "grey" fit. Normally the tie is broken by an ordinary random number. The watermark swaps that random source for a deterministic function of a secret key plus the recent words. The picks still look random to a reader, but they follow a hidden pattern.

How detection works

Anyone holding the key can score a passage and ask a simple question: do these word choices match what Claude would have picked using the key? If enough choices line up, the checker can assign a probability that Claude wrote the text. Anthropic says a watermark detection API is on the way, though implementation details are still being finalized.

Why readers cannot tell the difference

The watermark never pushes Claude toward strange words. It won't suddenly reach for "nubilous" in place of "grey". The model still chooses from the same shortlist it always would; only the tiebreaker changes. In internal testing, Anthropic reports no measurable effect on content, creativity, or readability.

The approach is a version of the SynthID-Text method published by Google DeepMind in a Nature paper in 2024. A live A/B test on a slice of Gemini traffic found no statistically significant difference in thumbs-up ratings between watermarked and unwatermarked outputs, and a controlled human study saw no side-by-side gap in quality. For users of models like, the change is a background layer they won't notice in day-to-day chats.

Where the Claude text watermark works well, and where it does not

The watermark thrives on long, open-ended writing where Claude makes many free choices. It struggles anywhere those choices dry up.

Two rules of thumb explain the table. First, the longer the passage, the more decisions Claude makes, and the more confidently a detector can call it. Second, the watermark only rides on choices where more than one word would be equally correct. If Claude writes "Isaac Newton's most famous work was called Principia…", the next word must be "Mathematica" and there's nothing for the watermark to act on. The same holds for "2 + 2 =": the answer is "4", and no nudge is applied.

Code inherits this constraint. Keywords, function names, and syntax usually have one right token, so the watermark stays out of the way. Free-form comments inside code are fair game, but they have negligible effect on the code itself.

What about editing and translation?

Light edits usually leave the pattern intact. A full rewrite that replaces every word will erase it, though at that point the text is arguably not AI-generated anymore. A translation produced by Claude does carry a watermark, because every word in the output was chosen by Claude.

Files and images use a different mechanism

Watermarking is text-only. When Claude produces a PNG, JPG, or SVG, it instead attaches a content credential in the file's metadata. According to C2PA, this is the same open standard camera makers and photo-editing apps use to record where an image came from. Nothing in the pixels changes; the credential is a cryptographically signed note stapled to the file. Anthropic will offer its own drop-and-check tool alongside any C2PA-aware reader.

Why now: the EU AI Act

Anthropic is rolling out the Claude text watermark to comply with the EU AI Act. In July 2026, Anthropic and around 190 other signatories signed the EU Code of Practice on Transparency of AI-Generated Content, which requires providers to mark AI-generated text. According to the European Commission, this transparency track is a central pillar of the AI Act's approach to synthetic media.

Because Anthropic doesn't yet have a clean way to scope the change by region, the watermark applies globally at launch. Models released before August 2, 2026 get a transition window and will be updated over the coming months. The company's growing policy footprint is part of a wider push: Anthropic recently named its first Chief Global Affairs Officer, as covered in. Other frontier labs are moving in the same direction, part of the regulatory wave discussed in.

What it means for users, privacy, and detection

The Claude text watermark is a signal about the model, not the person. It contains no user ID, no organization name, and no chat metadata. Nothing in the key would let anyone recover who prompted Claude or what the prompt said.

It also doesn't affect speed or cost. The watermark produces no extra tokens, so serving a watermarked model is the same price as serving an unwatermarked one. And it doesn't change ownership or authorship rights under Anthropic's terms.

One last point worth being clear about. A positive result only says "Claude was probably involved." It can't separate "Claude wrote this from scratch" from "Claude heavily edited this." It also can't detect text from other AI systems, because each provider uses its own key or a different scheme entirely. Public AI-detection services work differently again, hunting for stylistic tells like the overused "this isn't X, it's Y" construction or an unusual fondness for the word "quietly".

FAQs

Can I remove the Claude text watermark by editing the output?

Light edits usually leave the watermark intact enough for a detector to catch. A full rewrite where every word is replaced will remove it, but at that point the text has effectively become human-written. Anthropic says a watermark can only ever confirm that Claude was likely involved at some point.

Does the watermark identify me or my prompts?

No. The Claude text watermark applies to Claude's outputs, not to users. Nothing in the watermark or its key encodes your identity, your organization, or your conversation history with Claude.

Will the watermark slow Claude down or raise my API bill?

No. According to Anthropic, watermarking has a negligible effect on model speed and produces no extra tokens, so pricing is unchanged compared to unwatermarked serving.

Does watermarking work on Claude's code output?

Rarely for the code itself. Syntax and function names usually have one correct token, so the watermark cannot ride on them. It can appear in free-form comments, but that has no effect on how the code runs.

Is Claude the only model with watermarking?

No. Several major AI providers signed the same EU Code of Practice in July 2026 and are rolling out their own marking methods. Anthropic's approach is based on Google DeepMind's SynthID-Text, but each provider uses its own key and design choices.

When will a detection tool be publicly available?

Anthropic says a watermark detection API is coming, with implementation details still being worked out. There's no public release date yet, and the company has committed to sharing updates as the rollout progresses.