Building a text watermark remover for Claude, rewriting works best
Imaginary_Dinner2710 · reddit · 2026-08-23
The author details their technical journey building a watermark remover for Claude, experimenting on the SynthID Text scheme. They found that formatting cleaning and back-translation fail, while the DIPPER algorithm removes marks but corrupts facts 25% of the time. The most effective method is using an unmarked model to fully retell the text, achieving 100% removal and fact preservation. The author warns against verbatim copying during rewrites and provides open-source code and a live demo.
More from Safety
- Polymarket odds: Only 10% chance US enacts AI safety bill by end of 2026 — Polymarket · 2026-10-16
- 8 states sue OpenAI over alleged harm to children; Iowa AG says ChatGPT 'coached suicide' — Polymarket · 2026-10-16
- Reviewer will decline papers largely written by AI, citing author effort — IanArawjo · 2026-10-16
- Perplexity open-sources SPACE, the Firecracker microVM sandbox behind Perplexity Computer — perplexity_ai · 2026-10-16
- Banking regulators ramp up scrutiny of banks' AI use over stability fears — nordicinst · 2026-10-16
- Michigan professor Ben Green: why 'AI for Good' fails and self-regulation won't work — benzevgreen · 2026-10-16