Claude 'watermark' rumor debunked: random word lists flagged 100% AI by Pangram
rubenhassid · x · 2026-11-22
Ruben Hassid debunks the viral claim that Claude hides a "W-R-I-T-T-E-N B-Y C-L-A-U-D-E" acrostic watermark:
- Asked Claude for 50 random kitchen words and ran them through Pangram (an AI detector): labeled 100% AI-generated. Repeated with a fresh list — same result.
- 50 words written manually scored 100% human-written.
- Conclusion: no real evidence of a watermark; the detector signal has nothing to do with acrostics.
He also shares a workaround: paste a long Wikipedia-based prompt restricting Claude to only use sentences and words from it, and Claude's output gets flagged as 100% human-written.
More from Models
- Gemini subscriptions may work in third-party agent harnesses starting December 2 — testingcatalog · 2026-11-22
- Tesla FSD drives through fog so thick a human can barely see the road — XFreeze · 2026-11-22
- 'Gemini 5': a two-word post sparks speculation about Google's next model — inductionheads · 2026-11-22
- Screenshot reveals dual weekly + 5-hour rate limits on Gemini subscription quotas — lipeng0820 · 2026-11-22
- Meme asks if Qwen 4 27B will ever ship as Alibaba keeps delaying it — zyxciss · 2026-11-22
- OpenAI ships official system prompt templates to fix GPT-6.2's lazy, ask-too-many-questions behavior — lxfater · 2026-11-22