Gemini invisible character detector

Paste a passage taken from Gemini, or from a Google Doc or Gmail draft it contributed to, and this route reports the zero-width and non-printing Unicode present in it. The text is read, not rewritten, and the output is a description of observable signals rather than a verdict about anything.

The operational facts, since they constrain what the page can honestly claim: this is a server route. Text is posted to our AI server, the allowance is 5 runs per rolling 24 hours, and input is bounded at 1 to 12,000 characters. None of that is a browser operation, so nothing on this page claims local processing.

Most of the confusion around this query is a layer confusion. Google's published approach to text watermarking operates on token selection: as the model generates, candidate tokens are scored against context-derived keys, producing a statistical pattern detectable across a passage. The carrier is the word sequence.

An invisible character is a carrier at the byte level. The two do not overlap. A character detector has no visibility into token-selection statistics, and a character remover changes nothing about them, because deleting a zero-width space does not alter which words were chosen. Any page that advertises character stripping as Gemini watermark removal has collapsed two layers into one, and the resulting claim is false regardless of how well the stripper works.

Short procedure, with the reason for each step stated, because the reasons are where the errors are.

Step 1

Paste as plain text where the option exists. Pasting rich text wraps the content in formatting and changes what you are actually examining.

Step 2

Run the check and read the findings as evidence about a copy path, not about authorship. The distinction matters for what you conclude next.

Step 3

Treat removal as a separate decision. This route reports. The deterministic cleaner on this network's watermark routes removes, and it runs in the browser rather than on a server.

Coverage is a finite, checkable list, so it is published rather than implied.

Zero-width and joiners

U+200B, U+200C, U+200D, U+2060, U+FEFF. Width zero, non-empty in a byte comparison, which is precisely why equality checks fail on strings that render identically.

Space lookalikes

U+00A0, U+202F, U+200A. Distinct codepoints that render close enough to U+0020 that visual inspection cannot separate them. Cleaning substitutes an ordinary space.

Soft hyphen

U+00AD. Conditionally rendered: invisible until a line break falls inside the word, then a hyphen appears that was never authored.

Directional controls

U+202A, U+202B, U+202C, U+202D, U+202E. Embeddings, overrides and the terminating pop. An unbalanced one changes display order without changing the underlying sequence.

Explicitly not covered

U+200E, U+200F, U+061C, U+2028, U+2029, U+034F, U+FE00 to U+FE0F, U+2061 to U+2064, U+2066 to U+2069. Competing pages list these. The engine here does not implement them, so a clean result says nothing about them.

The deliberate-watermark hypothesis is the popular explanation and the weaker one. A marking scheme that survives a copy-paste would also have to survive a text editor's save, which these characters routinely do not, and it would be a strange design to hide a watermark in a carrier that any cleaner strips in one pass. The likelier account is that the training distribution contains them. A model has no keyboard, so no physical constraint filters out characters a person would never produce.

The second source is the rendering path, and it is empirically separable from the first. Chat interfaces emit HTML, and a copy operation from HTML carries whatever the layout used, commonly a no-break space or a hair space adjacent to punctuation. Google Docs and Gmail behave the same way. The diagnostic is repetition: the same character in the same position across every paste from one source indicates the path, not the model.

Everything that depends on byte equality while the human depends on appearance. Phrase search misses, find-and-replace skips rows, version control reports a diff on a line that looks unchanged, slug generation produces a URL containing a character no editor can select. The cost is not the bug, it is the time spent looking at two identical-seeming strings.

The security case is narrower and more serious. A zero-width carrier can hold arbitrary content inside a sentence that reads normally, which has been used to conceal payloads and to embed instructions intended for a model that later processes the document. Where untrusted text enters a system that acts on it, inspection is a control rather than a nicety.

No, and the inference fails in both directions. The presence of U+00A0 establishes that the text passed through a rendered surface, which describes Gemini, a news site, a PDF reader and a word processor equally well. Absence establishes even less, since a clean paste is trivially achievable. Treat the finding as evidence about transport, not about origin.

Nor does removal move a passage past a detector. Detection operates on lexical and structural statistics, which is the same layer the token-selection watermark occupies and a different layer from the one this page inspects. This route reports observable signals, makes no authorship claim, and predicts no third-party verdict.

Three usage patterns, distinguished by what failure they are trying to avoid.

Editorial and publishing

Inspecting a draft assembled from mixed sources before it reaches a CMS, where a hidden character becomes a permanent slug defect.

Engineering

Diagnosing a failed equality check or a rejected config value before rewriting the parser that was behaving correctly all along.

Intake and submission

Portals that reject input without diagnostics. A stray control character is a frequent cause and an invisible one.