Text Watermark Remover is a de-watermarking tool for Claude-generated text, and its official page is built around the purpose of "removing Claude watermarks," with the core approach being to split watermarks into two categories and handle them separately.
Purpose and Background The page states that Claude models released by Anthropic on or after August 2, 2026 embed invisible statistical watermarks in generated text, and that this is planned to extend to earlier models, in order to meet the EU AI Act's transparency requirements. This method is described as Anthropic's own solution, not Google's SynthID. Therefore, the tool's goal is to reduce the likelihood that content flagged as AI-written will be detected.
Main Capabilities: Two-Layer Processing
- Quick Clean: Scans and removes over 60 kinds of invisible Unicode characters locally in the browser in real time, including zero-width spaces, narrow no-break spaces, byte order marks, and more. The official description emphasizes that this step is 100% completed on the client side, does not upload text, and that detection and deletion are deterministic and can be verified by repeated scans, and that it is free and unlimited to use. At the same time, the page explains that line breaks, spaces, and tabs are formatting rather than watermarks, and will not be counted or removed.
- Deep Rewrite: For statistical watermarks, it holds that the signal is hidden in word choice, that there is no payload that can be deleted, and that the signal can only be diluted by rewriting the text with an LLM. It is officially explicitly labeled as best-effort, because Anthropic's official detector is not public, so thorough removal cannot be verified or promised. It supports up to 8,000 characters, and by default provides rewriting with "strong changes, readable output."
How to Use Users copy text from a Claude conversation and paste it into the tool; Quick Clean scans invisible Unicode characters in real time; "Quick Clean" is used to remove invisible characters, and "Deep Rewrite" is used to handle statistical watermarks, after which the cleaned result is copied. Rewrite history is kept only in the current browser and is not uploaded. Deep Rewrite requires signing in with Google to get 3 free rewrites for life, with the quota synced across devices; by sharing a personal link, users can get +3 additional rewrites when a friend tries it (up to 10 for life, once per day).
Use Cases The page lists two categories: first, text flagged as AI-generated by AI detectors (such as Turnitin) that needs to be handled before resubmission; second, publishing Claude-assisted content while hoping to reduce downstream risks brought by the spread of AI detectors.
Limitations and Notes The official description repeatedly emphasizes that "clean characters ≠ clean text," that Quick Clean cannot eliminate statistical watermarks, and that passing a scan does not mean a watermark does not exist. Statistical watermarks can only be probabilistically weakened through rewriting, and there is no public detection method to verify it. The page states that its own scanning is heuristic and is not a vendor detector. Regarding pricing, the page says Quick Clean is permanently free and unlimited, Deep Rewrite has a lifetime free quota, and paid plans start at $6 per month; the quota rules for login, sharing, and so on are as above. Questions such as whether it will lead to a Claude account being banned, whether removal is legal, and whether watermarks remain after copy-paste are raised in the FAQ format on the page, but the source material does not currently provide the specific answers.
Applicable Models and Text It is officially stated to be applicable to any Claude model, including Sonnet, Opus, Haiku, and newer models released afterward.
