AI Watermark Remover

An application that uses a multistep pipeline to remove any watermarks from AI generated text

What it does:

• takes text, strips the typographic markers LLMs usually leave behind (em dashes, zero-width characters, exotic spaces, curly quotes)
• runs it through a translation round trip (e.g. English -> intermediate language -> English), can select different translation provide (Google, Mymemory, Deepl)
• paraphrases the result with an LLM, keeping the meaning, quality and readability intact.

Paraphrasing runs through OpenRouter, OpenAI, Gemini, Claude or a local Ollama model.

Any step is optional, so you can configure watermark removing pipeline as you like.

Setup

uv sync
cp .env_example .env # then fill in the API key of the provider you use
cp config_example.yaml config.yaml # optional, the example is used until you do

Run

uv run streamlit run app.py

Paste the text into Input text field, then press Ctrl + Enter to confirm it, then press Process.

Configuration

• config.yaml - which steps to run, translation provider, intermediate language, paraphrase provider and model. It is gitignored so your settings stay yours; config_example.yaml is the version in the repository and is read as a fallback while config.yaml does not exist.
• .env - API keys (see .env_example). Only the key of the selected paraphrase provider is needed, ollama needs none. The google and mymemory translation providers are free and need no key; deepl needs TRANSLATOR_API_KEY.
• src/prompts.py - all prompts.

Every setting in config.yaml can also be overridden per run in the sidebar of the GUI. Those overrides live only in the browser session, until you press Save settings, which writes them back to config.yaml (comments included). Secrets are never written there, they stay in .env.

Results

watermark_detector/detect_watermark.ipynb takes one watermarked paragraph, runs it through each route of the pipeline and scores what comes back. The source text was generated by gemma-2b-it with a SynthID watermark applied under the public demo key set, so the detector can actually see it. Gemini's own keys are private, see watermark_detector/README.md.

metric	watermarked	en-de translated	en-de paraphrased	en-cn translated	en-cn paraphrased	en-ru translated	en-ru paraphrased	only paraphrased
score	1.0000	0.9998	0.0000	0.3261	0.0042	0.1861	0.0000	0.0000
z_score	+11.43	+5.68	+1.02	+4.10	+2.86	+4.19	+0.51	-1.18
avg g-value	0.5599	0.5299	0.5053	0.5215	0.5153	0.5217	0.5026	0.4940
is_watermarked	yes	yes	no	no	no	no	no	no

score is the Bayesian posterior the verdict comes from. z_score says how many standard deviations the mean g-value sits above the 0.5 that unwatermarked text gives, against a null measured at mean -0.09 and standard deviation 0.90.

Translation on its own does not remove the watermark. The German round trip comes back at 0.9998 and is still flagged. Chinese and Russian drop under the threshold but keep z-scores above 4, so there the watermark is degraded rather than gone.

Paraphrasing is the step that clears it, and on this sample it does not need the translation to help. Paraphrasing alone gives the cleanest result in the table: z = -1.18 is the only negative score, meaning no residual signal at all rather than a signal pushed under a threshold. Adding a round trip in front of it does not improve on that, and in the Chinese case leaves a trace at z = +2.86.

These are single samples per route, so treat the ordering as indicative rather than measured.

Layout

File	Purpose
app.py	Streamlit GUI
src/pipeline.py	Orchestrates clean -> translate -> paraphrase -> clean
src/cleaner.py	Symbol map, replacement and per-symbol statistics
src/translator.py	Round-trip translation with chunking
src/paraphraser.py	LangChain chain over the selected provider's model
src/prompts.py	Prompts
src/config.py	config.yaml + .env loading
watermark_detector/	SynthID watermark detection, optional extra, see its own README