A browser-based utility designed to identify and remove technical artifacts, hidden HTML, and statistical watermarks from Claude-generated text to ensure clean, human-readable prose.
What the product does and how it is positioned
Claude Watermark provides a suite of tools to identify and strip metadata and stylistic markers from AI-generated prose, focusing on deterministic findings like hidden Unicode characters and HTML class names.
The tool operates primarily within the user's browser to maintain privacy during the initial detection phase, cleaning text for various platforms by removing technical traces left behind when copying content.
Source-supported ways to use the product
The tool strips hidden HTML class names that are often included when copying text directly from the Claude web interface.
The rewrite engine processes sentences through alternative models to vary word choice and rhythm, helping to mitigate statistical watermarking signals.
The documented workflow, where available
Paste the text passage into the browser-based interface for immediate analysis of technical artifacts.
Examine the count and position of detected HTML tags, invisible characters, and stylistic markers surfaced by the checker.
Use the one-click removal tool for technical artifacts or initiate a rewrite pass to modify sentence rhythm and word choice.
Anthropic utilizes a statistical watermark based on keyed bias in word choice, which cannot be reliably detected without access to a private key. To address this, the tool employs a rewrite engine that processes every sentence through a different model.
This resampling technique is intended to disrupt the specific word-choice patterns that constitute the watermark. By varying sentence length and rhythm, the system aims to produce prose that lacks the predictable signals often associated with machine-generated output.
Checks to run with your own material and workflow
What was checked and when
Answers based on the source-checked product record
The tool removes hidden HTML class names, zero-width Unicode characters, non-breaking spaces, and stylistic markers like excessive em dashes.
Since statistical watermarks are embedded in word choice, the tool uses a rewrite pass to resample text through a different model to disrupt those patterns.
No, the initial check for artifacts and invisible characters runs entirely within the user's web browser without uploading data.
It strips the specific CSS class names that certain platform detectors use to identify text pasted directly from the Claude interface.
Text sent for rewriting is processed by a model provider and is not stored or used for training purposes after the task is complete.
