Claude’s Text Watermark Is Global by Default — Because Brussels Requires a Mark
Future Claude models will embed a SynthID-Text-style watermark — undetectable to readers, with no extra tokens or user identifiers — to comply with the EU AI Act’s August 2 marking rules, applied globally at launch because the lab cannot yet scope it by region. Watermarks do not prove authorship and die on rewrites.
Marking is the first EU AI Act obligation that actually touches every Claude sentence, not just GPAI eval paperwork. A global SynthID-style scheme is a bet that one key can satisfy Brussels without splitting the model. Limits are the story as much as the method: watermarks do not prove authorship, die on rewrites, and barely touch code — the workload enterprises actually buy.
Anthropic published How Claude’s text watermark works. Future Claude models will generate text containing a statistical watermark so anyone with Anthropic’s key can estimate whether Claude was involved. The lab uses a version of Google DeepMind’s SynthID-Text (Nature, 2024), in the family of schemes proposed by Scott Aaronson in 2022: it changes the source of randomness among near-tie next-token choices, not the meaning. Anthropic says there is no practical quality impact, no hidden characters, no extra tokens, no per-user identity, and no cost increase. The change is for EU AI Act transparency rules that applied from August 2, 2026; Anthropic and ~190 Code of Practice signatories must mark AI-generated content. Watermarking is global at launch because the company “doesn’t yet have a durable way to scope it by region.”
How it works (and where it fails)
| Claim | Anthropic’s wording |
|---|---|
| Method | SynthID-Text-style: keyed randomness over low-stakes word choices |
| Reader | Watermarked vs unwatermarked text is not distinguishable |
| Detection | Probability Claude was involved; not “human-written,” not “which other model” |
| Short / factual / code | Sparse or absent where there is only one correct next token; comments may still carry a mark |
| Edits | Light edits often survive; a full rewrite removes it |
| Files | Images/SVG get a C2PA content credential in metadata — not a text watermark |
| Older models | Transition period for models launched before Aug 2; roll-out “over the coming months” |
| Detector | A watermark detection API is “soon”; details still being worked out |
Not dual-filed as the Aug 2 statutory switch-on. Aug 14 is the lab implementation.
Limits
- Detection API not yet shipped.
- Watermarks fail on full rewrites and barely mark code.
- Whether educators treat “Claude-likely” as discipline evidence it cannot support is a misuse risk.