Claude watermark cleaner, not SynthID

A fixed list of 14 codepoints, applied deterministically. Eleven are deleted. Three are replaced with U+0020 because they occupy the visual role of a space and deleting them would concatenate the surrounding words. On top of that, runs of two or more spaces or tabs inside a line collapse to one, trailing whitespace is trimmed per line, and the output is trimmed at both ends.

The operation is local. Nothing is transmitted, no account exists, no allowance is metered and no character cap is imposed, because none of those are needed when the work happens in the page. The same input produces the same output every time, which is the property that makes the result checkable: scan the cleaned text and the count should be zero.

Anthropic's published account of Claude's text watermark describes it as a pattern in Claude's responses, undetectable to a reader but detectable to anyone holding the key that encodes it, and states directly that nothing is added to the text and there are no hidden characters. It is a version of the SynthID-Text approach published by Google DeepMind in a 2024 Nature paper, and it works by modifying the source of the randomness used to pick among words rather than the selection itself.

That places it categorically outside the reach of any character-removal operation. There is no token to delete. Anthropic's own note states that light editing probably will not remove the watermark and that a complete rewrite replacing every word will, and adds that the mark is sparser on factual passages and proofreading, where there are few word choices available to bias.

The invisible characters this route removes are a separate artefact with a separate cause. Both get called a Claude watermark, which is how a category of tools ended up advertising a removal it cannot perform. This page states the distinction rather than exploiting it.

Scan before cleaning. The report is the evidence that a change was warranted, and a zero result is as useful as a positive one.

Step 1

Paste the text. No cap applies, so a full file is acceptable. A sample loader is available for inspecting the report against known dirty input.

Step 2

Scan. Findings render inline with a per-item reason, and two independent counters report hidden codepoints and flagged phrases.

Step 3

Clean, read the badge for the removal count, then copy. Re-scanning the output verifies the operation, which is the advantage of a deterministic pass.

The set is fixed in this codebase, not configurable and not extended to make a larger number look impressive. Pages naming five characters and pages naming eighty-seven are both describing something else.

Zero-width, four entries

Zero-width space, non-joiner, joiner and word joiner. Each occupies a character slot while rendering nothing, so every exact-match operation sees a different string from the one on screen.

Byte-order mark

Meaningful at the head of a file and meaningless anywhere else. Mid-document copies break parsers, feeds and validators.

Soft hyphen

A conditional hyphen with no effect until the line wraps, at which point a hyphen appears in a word that never contained one.

Space look-alikes, three entries

No-break space, narrow no-break space, hair space. These are the substituted three, since removal would merge adjacent words.

Bidirectional controls, five entries

Two embeddings, a pop, and the left-to-right and right-to-left overrides. An override can make displayed order diverge from stored order, which is a security-relevant property in filenames and links.

The exception

A zero-width joiner flanked by pictographic characters composes an emoji sequence and is preserved. Blanket stripping of that codepoint breaks emoji, which is why the check exists.

A rewriting engine could restructure the prose and, by Anthropic's own account, a complete rewrite is what degrades the watermark. It is deliberately not what runs here. A deterministic pass over a named character set has properties a rewrite cannot offer: it preserves every word, it is verifiable by re-scanning, it cannot hallucinate, and it can be applied to source code without changing what the code does.

The trade is explicit. This route will never help with a statistical mark or with prose that reads as generated. What it will do is leave your text semantically identical and byte-clean, which is the requirement when the destination is a repository, a parser or a database rather than a reader.

Since the model adds none, they enter downstream. Rendered chat interfaces carry formatting characters that a selection copies along with the visible text. Rich-text clipboards preserve them where a plain-text field would not. Word processors and note applications insert no-break spaces around numerals and dashes as a typographic convenience.

Contamination tracks the copy route rather than the generator. The same response copied two ways yields two different scan results, which is why a scan beats an assumption about which assistant produced the text.

It does not remove Claude's text watermark, and no character tool does. It does not detect that watermark either, since detection of a keyed SynthID-Text mark requires the key. It makes no prediction about any AI detector's verdict and is not tuned against one.

It does not rewrite, paraphrase or humanize, so it changes no word choices at all. It does not normalise punctuation on this route, so curly quotes and dashes survive, and it does not collapse blank lines. There is no file input, no batch path and no API. Finally, removing a machine-readable mark does not extinguish a disclosure obligation under an academic, contractual or platform rule, and nothing on this page should be read as suggesting it does.

The result set divides into three. Humanizer products that rewrite and describe the output as undetectable. Tool pages that strip invisible characters while implying they address Anthropic's mark, in one case claiming in a single sentence to remove hidden characters so that the watermark becomes undetectable. And a small number of pages that draw the distinction correctly, one of which states outright that there is no character to strip and that no character-removal tool can remove the watermark, its own included.

The honest pages are also the ones with the tightest limits: a 500 character field and a minute-long cooldown between scans in one case. This route has neither, which is a consequence of doing the work locally rather than a claim of superiority over the analysis those pages perform.