Claude Watermarks Every Output Now. Detection Still Isn't Proof.
Every Claude model launched on or after August 2, 2026 embeds an invisible watermark in its text — globally, with no product surface able to opt out. The risk isn't the mark. It's the compliance gate teams are about to build on top of it.
NeuroX AI · August 17, 2026
Anthropic now embeds an imperceptible watermark in text from every Claude model launched on or after August 2, 2026 — the API, Claude Code, Cowork, Tag, and the third-party clouds. It applies at the model level, worldwide, and no product surface can opt out. The method is SynthID-Text, Google DeepMind's approach, whose Nature paper measured nearly 20 million Gemini responses and found "no change in LLM capabilities."
The mechanism explains where it works. The watermark biases the choice among words that are all equally valid — "overcast" or "grey" — so it needs a stream with real optionality to carry signal. Anthropic says so plainly: short passages are weak because "there are fewer word choices and thus less information to go on." Code, config, and structured output are the most constrained token streams your team generates. The output engineers ship most is the output the mark holds onto least.
The trap is what gets built on top. Anthropic states that a detected mark means Claude was "likely involved" and is "not fully conclusive" — and absence proves nothing, since editing, paraphrase, or a pre-August model all erase it. Any policy gate reading detection as "an AI wrote this" is a coin flip wearing a compliance badge.
Provenance you can defend is provenance you record: commit trails, review sign-off, eval logs. The watermark is a signal for the platform, not evidence for your auditor.