Remove Claude watermarks — the setup that actually works

Why your Claude text is suddenly watermarked 🏷️

The reason it's happening now is the EU AI Act's transparency rules, which came into force on August 2 and push the big labs to make AI-generated content identifiable. It now covers every current Claude model, plus older ones like Opus 4.7 and Sonnet 4.6 — see Anthropic's current model table for rollout details.

Google already does this with Gemini, Anthropic is rolling it out now, and OpenAI has signed the same EU code — so it's safe to assume everyone else follows. This is not a Claude problem you can dodge by switching tools. It's where the whole industry is heading.

The five-minute setup 🛠️

A developer called Guillaume Meyer published the free open-source watermarks-remover toolkit. For the quickest text workflow, start with its clean-user-facing-text skill. The fuller remove-ai-marks skill connects to the toolkit's cleaning service for text and files.

1. 1Open the GitHub project.
2. 2Click Code, then Download ZIP, and unzip it.
3. 3Open skills, then compress the clean-user-facing-text folder into its own ZIP. Keep that folder and its SKILL.md inside the ZIP.
4. 4In Claude, open Customize → Skills, choose Add or the upload option, and upload that ZIP.
5. 5Turn on code execution and file creation under Settings → Capabilities (on Team and Enterprise an owner turns it on in Organization settings), then start a new conversation with the skill enabled.

New to skills? My complete Claude guide covers how uploading and enabling them works. Then paste the text or attach your text file, and ask for the skill by name:

The full file toolkit 🧰

The current remove-ai-marks skill uses a separate service. Start it in an environment your agent can reach, then install the skill. The repository's installer builds the upload ZIP for Claude — use the current README for your platform. For a local setup with Python 3.10 or newer, run these from the downloaded repository folder:

• On Windows, use py instead of python3.
• Upload dist/remove-ai-marks.zip through Customize → Skills.
• Keep the local service running. A cloud task needs a service it can reach, with the service API key configured.

Once connected, ask it to use remove-ai-marks on your file. In a host that supports the installed slash command:

In Claude Code, the project's plugin installs with:

Then use /watermarks-remover:remove-ai-marks. The full skill still needs the service. Read its inspection report and keep a copy of the original file before accepting the cleaned output.

Light pass or heavy pass? ⚖️

For prose, the tool will offer you the rewrite, and this is where you make a judgement call. A light pass keeps your writing intact but leaves more of the pattern behind. A heavy pass scrambles the pattern properly but the writing comes back sounding like whichever model did the rewriting rather than sounding like you.

This is the light paraphrase prompt the skill uses:

And the stronger one, which reads better but drifts further from what you originally wrote:

What the watermark actually is 🔬

When Claude writes a sentence there are usually several words that would fit equally well in any given spot, and the model is picking between them based on probability. The watermark works by nudging that choice using a secret key, so across a long enough passage the pattern of which near-equivalent words got picked forms a signal that a detector holding the key can spot.

The text reads completely normally and no individual word looks unusual, because the signal only exists in the pattern across hundreds of words.

It works at the model level rather than inside any one app — once a model has it, the mark is there whether the text came out of the Claude app, the API, Claude Code, Cowork, or Claude in Slack. Anthropic also attaches signed provenance metadata (C2PA content credentials) to files Claude creates, which anyone can check with its free Content Checker tool.

What actually weakens it:

• Fixing a typo or tweaking a few words: barely touched. Most of the original word choices are still sitting there.
• Changing headings and formatting: no effect. The signal is in the prose, not the layout.
• Rewriting heavily in your own words: genuinely disrupted. You're replacing the choices that carried the pattern.
• Short text, a few sentences: hard to read either way. Not enough words for the pattern to be reliable.
• Code: weak to begin with. Little room to nudge word choice, and formatters strip the rest.

What the tool actually cleans 🧹

It works in three layers, and they are not equally reliable:

• Layer A — invisible characters. Zero-width spaces, narrow spaces, bidirectional marks and other odd Unicode that AI tools slip in. Verifiable — it counts exactly what it pulled out and shows you.
• Layer B — the statistical watermark. The word-choice pattern, attacked by getting a second AI to rewrite your text. Best effort. Nobody outside the vendors can currently confirm it worked.
• Layer C — files. C2PA content credentials, EXIF and XMP metadata in PNG, JPEG, SVG, PDF, DOCX, ODT, HTML and MD. Verifiable — it reports each field it stripped.

Even though the new watermark isn't made of invisible characters, AI tools do still leave those characters in your text, and they're a strong tell on their own. They also break things — paste one into a CMS or a code editor and you get formatting bugs you'll spend an hour hunting.

Where people get this wrong ⚠️

• Missing the setup step. The self-contained text skill needs code execution. The full remove-ai-marks skill needs a running, reachable cleaning service. Check which skill you installed and start a fresh conversation after enabling it.
• Rewriting Claude text with Claude. You're handing the job to the thing you're trying to get away from.
• Skipping past the inspection report. That report is the only genuinely verifiable output in the whole process.
• Running a light rewrite and calling it done. A gentle pass that keeps your voice also leaves most of the original word choices sitting there. You don't get both.
• Forgetting to clean again afterwards. The rewriting model is an AI as well. Clean, rewrite, then clean again.
• Forgetting that images and documents carry their own marks. People clean the text and then publish the screenshot next to it with full content credentials and camera metadata attached.

Links 🔗

• watermarks-remover (GitHub)— The free open-source toolkit
• Anthropic's model marking table— Which models watermark and how
• Humanizer skill guide— Make the cleaned text read human too
• The complete Claude guide— Skills, setup and everything else