What is this?
An experimental Python CLI that tries to strip Claude's statistical text watermark: it removes suspicious invisible Unicode characters, then has a non-Claude model rewrite the text. It doubles as a critical demo arguing such watermarks are easy to circumvent.
An experimental Python CLI that tries to strip Claude's statistical text watermark: it removes suspicious invisible Unicode characters, then has a non-Claude model rewrite the text. It doubles as a critical demo arguing such watermarks are easy to circumvent.Is it a library I import, or a program I run?A single command-line script, cleanclaudewatermark.py, written in pure Python — it relies only on the standard library, so there are no extra packages to install.
A single command-line script, cleanclaudewatermark.py, written in pure Python — it relies only on the standard library, so there are no extra packages to install.How do I run it on my text?Point it at a text file or run it on the macOS clipboard directly; you can also force a specific local Ollama model. By default it makes up to two rewrite attempts and keeps the stronger one, stopping early once 70% of five-token sequences are already perturbed.
Point it at a text file or run it on the macOS clipboard directly; you can also force a specific local Ollama model. By default it makes up to two rewrite attempts and keeps the stronger one, stopping early once 70% of five-token sequences are already perturbed.What do I need installed besides Python?A rewrite engine: a local model via Ollama is used first, with the Codex CLI already configured on your machine as a fallback. The script itself has zero external Python dependencies.
A rewrite engine: a local model via Ollama is used first, with the Codex CLI already configured on your machine as a fallback. The script itself has zero external Python dependencies.Can it actually guarantee the watermark is gone?No — real success can't be measured until Anthropic publishes its detector and thresholds. Rewriting can also subtly change meaning, so human review is still advised for legal, medical or financial text.
No — real success can't be measured until Anthropic publishes its detector and thresholds. Rewriting can also subtly change meaning, so human review is still advised for legal, medical or financial text.Does my text leave my machine?With the default Ollama engine everything stays fully local. In Codex fallback mode the text is sent to whatever service your Codex CLI is configured with, so the README says to prefer local Ollama for confidential documents.
With the default Ollama engine everything stays fully local. In Codex fallback mode the text is sent to whatever service your Codex CLI is configured with, so the README says to prefer local Ollama for confidential documents.Why does the author think Claude's watermark is a bad idea?Thesis: a mark you can weaken by rephrasing is not reliable proof of origin, so it mainly incentivizes a second AI pass — more compute, latency, cost and energy for the same content, plus ambiguous signals when someone uses Claude to edit or translate their own text.
Thesis: a mark you can weaken by rephrasing is not reliable proof of origin, so it mainly incentivizes a second AI pass — more compute, latency, cost and energy for the same content, plus ambiguous signals when someone uses Claude to edit or translate their own text.How is the watermark believed to work?Per the README, it is not hidden Unicode: it is applied during generation at the model level and survives copy-paste, printing and OCR. The most credible hypothesis — an inference, not confirmed by Anthropic — is a slight statistical bias toward certain tokens, similar to SynthID Text.
Per the README, it is not hidden Unicode: it is applied during generation at the model level and survives copy-paste, printing and OCR. The most credible hypothesis — an inference, not confirmed by Anthropic — is a slight statistical bias toward certain tokens, similar to SynthID Text.What does the script protect from rewriting?Code blocks, inline code, URLs and variables are preserved, and the script refuses to return output if any number has changed during the rewrite.
Code blocks, inline code, URLs and variables are preserved, and the script refuses to return output if any number has changed during the rewrite.How do I know if the cleanup actually worked?The script prints the engine used, the count of invisible characters removed and the lexical variation rate to stderr; a five-token sequence variation under 60% triggers a warning.
The script prints the engine used, the count of invisible characters removed and the lexical variation rate to stderr; a five-token sequence variation under 60% triggers a warning.What are similar projects?Baize, brigade, aichat, vomit, meetily, im-not-ai, opencodex, open-codesign
Baize, brigade, aichat, vomit, meetily, im-not-ai, opencodex, open-codesignWhich topics is it in?transparency, utilities, text-processing, privacy-tools, ai-tools, local-llm, ollama, anthropic, claude, codex, privacy, llm
transparency, utilities, text-processing, privacy-tools, ai-tools, local-llm, ollama, anthropic, claude, codex, privacy, llm