Anthropic Starts Watermarking Claude Text Output Worldwide

Anthropic announced on 14 August 2026 that it has implemented text watermarking across all Claude models released after 2 August. The technique originates in work Google DeepMind published in Nature in 2024.

How it works

There is no visible mark and no appended disclaimer. The signal is in the word choices themselves. When Claude generates text it repeatedly reaches points where several words would serve equally well; a cryptographic key biases which one gets selected. The resulting distribution is statistically detectable without being perceptible to a reader.

Detection is statistical, so it weakens on short passages and on text that has been heavily edited or paraphrased.

Why now

EU AI Act Article 50 transparency obligations took effect on 2 August 2026. Anthropic signed the EU Code of Practice on AI-Generated Content in July. Non-compliance can draw fines of up to EUR 15 million or 3% of total global annual turnover, whichever is higher.

The rollout went live globally rather than region by region - Anthropic said it could not implement regional controls at launch. Older Claude models have until 2 December 2026 to be covered.

For anyone publishing Claude-assisted text, the practical change is that the output now carries a detectable provenance signal by default, worldwide, regardless of whether the reader is in the EU.