Claude now watermarks all output text: user reproduces SynthID and open-sources a removal workflow
Imaginary_Dinner2710 · reddit · 2026-09-07
A Reddit user found that Claude (Fable 5.1) now embeds a statistical watermark in all generated text, including translations, based on Google DeepMind's SynthID method — an article written by the author and merely translated via Claude comes back watermarked. Meanwhile, the promised detector is only in private preview for select organizations, so regular users can't check their own text before publishing.
Reproducing SynthID with a self-made test key, the author found:
- A full rewrite removes the watermark but introduces factual errors
- Back-translation (even via Chinese) does not destroy the mark
- Simply asking another model (e.g. Qwen) to rewrite leaves substantial signal intact
The working approach: multiple iterative runs, each time highlighting unchanged 5-gram sequences to guide rephrasing, then fact-checking and fixing errors. The author open-sourced the setup and published a free demo, while cautioning it can't guarantee removal of production watermarks. They also note the mark can't distinguish AI-generated from AI-edited text, disproportionately affecting students, scientists, SEO specialists and non-native English speakers.
More from Safety
- If an agent can rewrite itself, does your consent still mean anything? — tallmetommy · 2026-10-26
- 109 of 529 AI Agent Incidents Involved No Adversary, Arguing Red Teams Must Test Normal Operation — anacondainc · 2026-10-26
- Cryptographer Matthew Green: if your code looks vulnerable, fix it now — matthew_d_green · 2026-10-26
- Filtering secrets on the write path: why agent memory deserves more suspicion at ingest than retrieval — Classic-Being-7960 · 2026-10-26
- After OpenCode key leak: 5 practical rules for keeping keys away from agents — jacobholm_dev · 2026-10-26
- Formal verification spat: specs are inherently limited against real-world threats — davidmanheim · 2026-10-26