Claude Watermarks: Anthropic Details SynthID Detection, Editing Limits and Code Impact
Anthropic explains Claude's SynthID-Text watermark, detection API, editing limits and lighter code marking as it prepares for EU AI Act rules amid backlash.
Summary
On August 15, 2026, TechCrunch reported that Anthropic had published a Friday blog post explaining Claude’s planned text watermark, introduced to comply with the EU AI Act’s Transparency Code requirement for identifiable AI content. Claude will use SynthID-Text, outlined by Google DeepMind in 2024, to encode a key-detectable pattern through low-stakes word choices such as “overcast” or “grey.” Anthropic says readers will notice no difference in output or quality, and it plans a detection API. Unlike Pangram and similar AI detectors, which analyze stylistic tells, the API will check for an embedded watermark.
Light editing probably will not remove the pattern completely, but replacing every word will, although Anthropic argues the rewritten result may no longer qualify as AI-generated. Detection in human text proofread or edited by Claude will depend on length and editing intensity, with little or nothing to mark when nearly all words remain human-written. Working code offers fewer arbitrary choices, so its watermark will be weaker, although comments and interchangeable terms can carry it with negligible effect on the code. The policy drew polarized Reddit reactions, while Business Insider reported that dozens of X users claimed to cancel Claude subscriptions. Anthropic says other major model developers signed the same Code of Practice and will implement their own watermarks.
Positives
- SynthID-Text, outlined by Google DeepMind in 2024, gives Claude a key-detectable watermark while Anthropic says output quality and appearance remain unchanged.
- Anthropic plans a watermark detection API that checks embedded patterns instead of relying on stylistic clues used by services such as Pangram.
- Light editing probably will not completely remove Claude’s watermark, providing some resilience against modest revisions.
- Code comments and interchangeable terms can carry watermarks with negligible effect on the working code Anthropic says Claude will produce.
- Other major model developers signed the same Code of Practice and, according to Anthropic, will introduce their own watermarks.
Risks & concerns
- Business Insider reported that dozens of X users claimed to cancel Claude subscriptions after Anthropic announced the watermarking policy.
- A complete rewrite replacing every word will remove Claude’s watermark, limiting its ability to trace heavily transformed output.
- Proofread or edited human text may have little or no detectable watermark, depending on its length and how extensively Claude changed it.
- Working code will carry less watermarking because functional requirements leave Claude fewer equally valid word or term choices.
- Reddit reactions included claims that watermarking targets innocent Claude users, reflecting distrust despite its EU AI Act compliance purpose.