AI watermark remover, exact set

A closed set of 14 codepoints, a defined action for each, and no behaviour outside that set. Eleven are deleted. Three are substituted with U+0020, the ordinary space. Every other byte in the input appears unchanged in the output, in the same order.

That specification is the product. A tool that cannot tell you precisely which characters it acts on is asking you to accept an unbounded edit to a document you are about to publish, and the category is full of them. Everything this AI watermark remover does is enumerable, and it is enumerated below.

Because the action has to preserve meaning, and a uniform rule cannot. The no-break space U+00A0, the narrow no-break space U+202F and the hair space U+200A all occupy the position of a word boundary. Delete one and the tokens on either side concatenate, so a cleanup silently corrupts the text it was run to repair.

The other eleven occupy no such position. A zero-width space sits inside or beside a token without separating it, so removing it restores the string a reader believes they are looking at. The distinction is recorded per entry rather than applied by rule, which is why the replacement behaviour is stated as 11 and 3 rather than as a single number.

The operation is local, unmetered and unbounded in length, so the only constraints are procedural.

Step 1

Supply the artefact unmodified. Any intermediate plain-text buffer has already performed a lossy version of this operation, so the input no longer represents the source.

Step 2

Read the per-codepoint counts first. The distribution identifies the upstream cause: scattered no-break spaces suggest a word processor, a leading byte-order mark suggests a file encoding, dense zero-width spaces suggest generated output.

Step 3

Verify at the destination, not here. Because the removed characters were invisible, the only observable confirmation is downstream behaviour: a slug that resolves, an import that parses, a character count that matches the visible length.

Fourteen entries, each with a defined action. Nothing outside this list is touched by this route under any circumstances.

Deleted: U+200B, U+200C, U+200D

Zero-width space, non-joiner and joiner. The non-joiner is semantically significant in Persian and several Indic scripts, where removal is a content change rather than a cleanup.

Deleted: U+2060, U+FEFF

Word joiner and byte-order mark. A byte-order mark occurring anywhere other than position zero is a concatenation artefact and breaks parsers, feeds and validators.

Deleted: U+00AD

Soft hyphen. Invisible until reflow, at which point it renders a hyphen in a token that does not contain one.

Deleted: U+202A to U+202E

The bidirectional embedding, pop and override controls. An unbalanced override alters display order, so a filename or link can render differently from its actual value.

Substituted with U+0020: U+00A0, U+202F, U+200A

No-break space, narrow no-break space, hair space. Word-boundary positions, so substitution rather than deletion.

Excluded: U+2014, curly quotes, U+180E

The em dash and quotation marks are visible glyphs and out of scope. The Mongolian vowel separator appears in other published lists and is absent from ours, which is stated rather than obscured.

The operation is deterministic: identical input yields identical output, with no model, threshold or sampling involved. It is also idempotent. Substitution emits U+0020, which is not a member of the removal set, so a second pass over cleaned text finds nothing and changes nothing. Running it twice is safe and running it a hundred times is the same as running it once.

The awkward property is verification. The characters removed were invisible, so the output looks identical to the input and success is not directly observable. This is why the counts matter more than the cleaned text does: the report of what was found is the only evidence the pass did anything, and the confirmation proper happens in whatever system was failing before.

Guides in this category publish a character class along the lines of a range from U+200B to U+200D plus a few singletons, to be applied with a global replace in an editor, in Python or in Node. It is close to right and it contains one defect.

The range spans U+200D, the zero-width joiner, and that codepoint has a second job: placed between two pictographic characters it composes them into a single glyph. A blanket replace therefore decomposes every combined emoji in the document into its constituent parts, silently, in text the author believed was only being cleaned. Our implementation inspects the codepoints on both sides of each joiner and passes over the ones performing composition. If you prefer to do this by hand, exclude U+200D from the class and handle it separately.

Every character in the set exists because some writing system or typographic convention needs it, so blanket automation over unexamined text will eventually destroy something intentional.

The clearest case is the zero-width non-joiner in Persian and several Indic scripts, where it is orthography rather than noise. The no-break space is conventional around French punctuation and between a value and its unit. A soft hyphen in a narrow measure may be deliberate hyphenation control. In each case the removal is still performed exactly as specified, which is precisely the problem: the tool is correct and the decision to run it was not.

Not materially. Detection systems weigh lexical and structural properties of the prose, and encoding-level characters are not among the features they consider. The most technically careful guide in this result set states that stripping invisible characters is necessary but not sufficient, and names perplexity, burstiness and vocabulary as what removal cannot touch. We have no evidence to offer against that and do not contradict it.

Other tools on this term advertise bypassing specific detection services by name. We hold no affiliation with any of them, cannot observe their outputs, and make no prediction. The returns this operation actually delivers are mechanical and verifiable at your own desk, and that is the entire claim.