Home › How it works
How the Claude watermark remover works
Text mode is live: it removes invisible Unicode characters and typography artefacts from pasted text in your browser and reports every change by code point. Image mode is not yet, though the detection and masking sequence is. Neither mode touches provenance signals.
Text mode: what it scans for
Real characters that are genuinely present in AI output and genuinely break things: zero width space, zero width joiner and non joiner, soft hyphens, word joiners, non breaking spaces and byte order marks, plus em dashes, smart quotes and ellipses.
These break git diffs, database keys, CSV parsing and search indexes. Removing them is text hygiene. It is not watermark removal, because Anthropic’s text watermark is statistical and no character carries it.
- What Claude’s watermark actually is, and why a cleaner cannot touch it
- Every character it looks for, with the failure each one causes
Image mode: how a watermark is removed
The pixels under a mark were overwritten when the file was exported, so removal is reconstruction rather than undo. Detect the marked region, estimate its opacity, mask with a feathered edge, then inpaint from the surrounding texture and colour match the patch.
- The four stages in full, plus which images come out clean
- Step by step on a real photo
What neither mode does
Text mode does not defeat Anthropic’s statistical watermark. Image mode does not strip C2PA content credentials. Both limits are deliberate and permanent. See the Acceptable Use Policy.