TechCrunch: how Anthropic's text watermarks work — and what they mean for writers

On August 15, 2026, TechCrunch published a detailed follow-up to Anthropic’s earlier watermarking announcement, explaining how the system works at a technical level and what the practical effects are for people who use Claude as part of a writing workflow.

The system uses SynthID-Text, an approach developed by Google DeepMind in 2024. The watermark works by making low-stakes choices between synonymous words during text generation, embedding a statistically detectable pattern across the output. The pattern is not visible to readers — the text reads the same — but it is detectable through a key that Anthropic plans to release as a detection API. Crucially, this approach differs from content analysis-based AI detectors, which try to infer authorship from stylistic features like sentence rhythm or phrase patterns. SynthID-Text embeds the signal at generation time; detection does not depend on stylistic inference.

The resilience of the watermark to editing is the detail most relevant to writers using Claude professionally. Light editing — rearranging sentences, adding context, changing examples — does not remove the watermark. A complete rewrite, where every word is replaced, does. Anthropic acknowledges the obvious inference: if every word has been replaced, the question of whether the text is still AI-generated is not straightforward.

The watermark is active on all Claude models released after August 2, 2026, and Anthropic has indicated it will extend support to older models. It applies across Claude’s consumer and professional surfaces: the API, claude.ai, Claude Code, and Claude Cowork. The EU AI Act’s Transparency Code, which also took effect on August 2, is cited as the regulatory context.

For professional writers who use Claude as a drafting or research tool, the practical implication depends on their workflow. Writers who produce AI-assisted drafts and then substantially revise them — changing wording throughout, restructuring, adding original analysis — are likely to reduce or remove the watermark through that editing process. Writers who submit lightly modified AI output are more likely to retain detectable markers.