Anthropic is adding invisible watermarks to text produced by Claude, but you won’t find a hidden tag, special character or piece of metadata buried inside your next AI-generated paragraph. Instead, the watermark is created through the words Claude chooses while generating a response.
The technology is based on SynthID-Text, a technique developed by Google DeepMind and published in Nature in 2024. Its basic idea is surprisingly straightforward: language models regularly encounter situations where several different words could naturally come next. Claude’s watermarking system subtly influences those choices according to a statistical pattern.
Take a simple weather description. Claude might reasonably describe a day as “grey,” “cloudy” or “overcast.” With watermarking enabled, the model can be nudged toward one of those equally appropriate options. One individual choice reveals essentially nothing. Across hundreds or thousands of words, however, those choices can form a detectable pattern.
That pattern is the watermark.
Importantly, Claude isn’t inserting invisible Unicode characters or attaching an identifier to every response. Anthropic says the watermark contains no information that can identify a particular user, company or conversation. It also doesn’t require extra tokens and should have negligible impact on how quickly Claude responds.
Detecting the watermark requires analyzing the text with the appropriate key. Anthropic plans to provide a detection API that can estimate the likelihood that a passage was generated or processed by Claude. This is different from conventional AI detectors, which typically examine characteristics of writing and attempt to determine whether the style resembles machine-generated text.
There is an important limitation: the result isn’t proof.
Watermark detection can indicate that Claude was probably involved with a piece of text, but it cannot conclusively establish authorship. It also cannot necessarily tell the difference between something written entirely by Claude and human text that Claude substantially rewrote.
The system becomes more reliable as the amount of generated text increases. Longer articles give Claude more opportunities to make the discretionary word choices needed to build a detectable statistical signal. Very short responses provide fewer clues.
Editing complicates things further. Changing a handful of words may leave enough of the original pattern intact for detection. Rewriting everything can effectively remove it. Similarly, asking Claude to make only minor corrections to human-written material may not introduce enough watermarked choices for a confident result.
Code is another difficult area because programming syntax often gives a model fewer legitimate choices than ordinary prose. Comments and descriptive text can still carry a watermark, while translations are better suited to the technique because Claude is choosing most of the resulting words.
For images and other supported files, Anthropic will use C2PA content credentials in metadata instead of the text-based system.
Anthropic is rolling the watermark out globally as part of its response to transparency requirements under the EU AI Act, with older Claude models expected to receive support during the transition period. Other major AI developers that signed the EU Code of Practice are expected to introduce their own approaches as well.
The result isn’t an indestructible fingerprint for AI writing. It is closer to a statistical trail—one that becomes easier to see as Claude writes more, but can fade when humans substantially rewrite what the model produced.
