Free Gemini Watermark Remover & Checker for Text & Code
Paste Gemini-generated text or code, see every hidden watermark character exposed in an x-ray view, and copy a clean version in one click. Free, online, no sign-up, and nothing you paste ever leaves your browser.
More: what was found, x-ray view and options
X-ray view
Your text with every hidden character exposed as a labelled chip. Hover a chip for its Unicode name.
What was found
| Code point | Character | Count | Action |
|---|
Does Gemini watermark the text it generates?
Yes. Google applies SynthID watermarking to text as well as images, audio and video generated by its models. For text the mark is embedded in token selection, so it survives copying and light editing, and removing hidden characters has no effect on it.
Last verified: 11 August 2026
SynthID is the most broadly deployed watermarking system of any vendor, because Google applies it across four modalities rather than one. For text specifically, the important consequence is that the signal is part of the output itself, not an attachment to a file, so it travels with the words.
How SynthID works on text
Language models produce a probability distribution over the next token and sample from it. SynthID for text intervenes in that sampling step. Using a key, it derives a pseudorandom score for candidate tokens at each position and biases the choice toward higher-scoring ones, within the range of continuations the model already considered acceptable.
The result is text a reader cannot distinguish from unwatermarked output, because every word chosen was already a word the model was willing to choose. What changes is the aggregate: across a long enough passage, the average score of the tokens actually selected sits measurably above what chance would produce. A detector holding the key computes that average and reports a confidence.
Two properties of this approach shape everything else about it:
- It degrades smoothly. Paraphrasing part of a passage removes part of the evidence. There is no threshold where the watermark abruptly stops existing, and no single edit that deletes it.
- It needs entropy. Where the model has essentially no choice, a direct quotation, a formula, a well-known list, there is nothing to bias, so those spans carry little or no signal. Very short outputs and highly constrained output carry less.
SynthID across modalities
| Output | Where the mark lives | What removes it |
|---|---|---|
| Text | Token selection across the passage | Substantial rewriting only |
| Images | Imperceptible modification of pixel data | Heavy crops, re-generation, severe re-compression |
| Audio | The spectrogram, converted back to waveform | Aggressive re-encoding and editing |
| Video | Frame content | Heavy re-encoding and re-editing |
None of these are metadata. That is the design goal: metadata is stripped by any platform that re-encodes an upload, whereas a watermark embedded in the content itself survives the ordinary journey of a file through the internet. Google's SynthID documentation, published by Google DeepMind, is the primary source for the mechanism, and it is where we check before revising this page.
Detection is not open to everyone. Verification requires the key. Google exposes detection through its own tooling rather than as a public library, and there is no way for a third-party website, including this one, to check text for SynthID. Anything claiming to do so in your browser is not doing it.
What this tool can and cannot do for Gemini text
Can
- Find hidden Unicode characters in text copied out of Gemini or a Google Workspace document, which in practice is mostly no-break spaces and thin spaces from the rendered page.
- Show every one of them in place in the x-ray view, with its code point and Unicode name.
- Return a cleaned copy safe for spreadsheets, databases, code and forms.
Cannot
- Detect SynthID. Detection requires Google's key, and no browser-side tool has it.
- Remove or weaken SynthID. The mark is in which words were chosen; deleting invisible characters changes no word.
- Tell you anything about images, audio or video. This tool reads pasted text only.
A practical note for Workspace users
Text that has passed through Google Docs on its way out of Gemini tends to accumulate more invisible characters than text copied straight from the chat interface. Docs uses non-breaking spaces for certain layouts, keeps soft hyphens from imported documents, and exports line separators rather than newlines in some paths. If you are assembling content through Docs, run the result through the checker before it reaches a system that cares about exact strings.
Questions about Gemini and SynthID
Can SynthID be removed from text by cleaning it?
No. SynthID for text is embedded in which tokens the model selected, not in characters added to the output. A cleaner that removes invisible characters leaves every token in place, so the signal is unchanged. Only rewriting enough of the passage to replace the model's word choices degrades it.
Does SynthID survive copy and paste?
Yes, for text. Because the mark is carried by the words themselves rather than by file metadata, copying the text into another application takes the watermark with it. That is precisely the advantage a watermark has over metadata-based provenance.
Can I check my own text for SynthID?
Not with this tool or any other third-party tool. Detection needs Google's key and is available through Google's own tooling. Any site offering browser-based SynthID detection is not performing a real check.
Does a short Gemini answer carry the watermark?
It carries less of it. The signal accumulates over token decisions, so a one-sentence reply provides far fewer data points than several paragraphs, and detection confidence is correspondingly lower. Highly constrained output, such as a direct quotation or a formula, carries little signal because the model had no real choice to bias.
Is SynthID the same as the C2PA content credentials on images?
No, though Google uses both. C2PA is signed metadata attached to a file and is lost when the file is re-encoded or the content is copied out. SynthID is embedded in the content and survives that. They are complementary layers, not alternatives.
Primary sources
- SynthID at Google DeepMind - the provider's own description of the watermarking family.
- Regulation (EU) 2024/1689, the AI Act, whose Article 50 sets the transparency obligation.
Related reading
- Claude's confirmed watermark - the second major deployment of the same class of mark.
- What Perplexity answers actually carry - where the underlying model decides the answer.
- Hidden characters in text - the class of problem this checker does solve.
- Statistical versus character watermarks - why one survives cleaning and the other does not.