How Anthropic's Invisible Claude Text Watermark Will Work

Anthropic has published its clearest explanation yet of how it will watermark the text Claude generates, answering the technical questions that have swirled since the company revealed earlier this week that it would comply with the EU AI Act's Transparency Code. The short version: Claude will bias inconsequential word choices to embed a pattern that is invisible to readers but detectable by anyone holding a secret key.

How the watermark works

Rather than attaching metadata to the end of a response, Anthropic is adopting the SynthID-Text approach first outlined by Google DeepMind in 2024. When Claude makes "low-stakes" decisions, such as choosing between "overcast" and "grey" to describe the weather, it biases those choices according to a key held by Anthropic. The result is a statistical signature woven into the text itself, not a flag bolted onto it.

"Watermarking does not impact the quality of Claude's output," the company said. "To a reader, a watermarked response is indistinguishable from an unwatermarked one."

Anthropic also confirmed it plans to release a watermark detection API, and drew a sharp line between watermarking and the probabilistic AI detectors sold by companies like Pangram, which look for stylistic "tells" in writing. Checking for a watermark, the company stressed, is "fundamentally different" from guessing whether a model wrote something.

Editing and code

The watermark is resilient but not unbreakable. Light editing "probably won't remove the watermark completely," Anthropic said, while a full rewrite that replaces every word will. For text that Claude only proofread or lightly edited, there is "very little (if anything) for the watermark to attach to," because nearly all the words were written by a human.

Code gets a lighter touch. Because Claude must produce working code, it has less freedom to make arbitrary word choices, so the watermark appears mainly in comments and other places where genuine choice exists, with a "negligible effect on the actual code produced."

Why it matters

Anthropic framed the change as industry-wide rather than a solo move, noting that "other major model developers have signed the same Code of Practice and will be implementing their own watermarks." That makes this a preview of the default trust infrastructure builders will soon inherit: output provenance is becoming a platform feature, a detection API is on the way, and code generation, the workload that matters most to automation teams, will be far less affected than prose. The disclosure lands a day after Google made visible watermarks optional on its own AI generations while keeping invisible SynthID provenance intact.