Two meanings of “AI watermark”
People use the phrase for both hidden formatting characters and statistical generation signals. They are not interchangeable.
| Type | Where it lives | Can a text cleaner remove it? |
|---|---|---|
| Invisible Unicode | Literal code points in copied text | Yes, when those characters are identified |
| Statistical watermark | A pattern across token choices | No, not by deleting hidden characters |
| Metadata | A file or document container | Sometimes, but not from plain pasted text |
A zero-width space is a real character. It has a Unicode code point even though it occupies no visible width. A statistical watermark is different: every displayed character can be ordinary, while the sequence still carries a detectable pattern.
How SynthID changes generated text
Language models generate text one token at a time. At each step, the model assigns probabilities to possible next tokens. Google DeepMind explains that SynthID adds information by modulating those probabilities during generation.
The resulting word choices create a pattern that a detector can score. Nothing needs to be appended to the final text, and there may be no unusual Unicode characters to find.
Key point: SynthID is applied while a compatible model generates text. It is not a post-processing stamp that a browser can simply delete.
The open-source reference implementation uses a watermark configuration containing keys, n-gram length, and sampling parameters. Different configurations create different watermarks. Google’s repository also notes that its reference classes are research examples rather than a production service.
Why generic detection is difficult
The reference implementation offers mean, weighted-mean, and Bayesian scoring approaches. But detection is tied to the watermark configuration. The Bayesian detector must be trained for each unique watermarking key on representative watermarked and unwatermarked data.
Text length matters too. Google reports that SynthID works best on longer, varied responses. A short sentence contains less evidence than an essay. A score is therefore a likelihood under a particular detector, not universal proof of authorship.
What Unmark Text removes
Unmark Text audits literal formatting artifacts: zero-width spaces and joiners, bidirectional controls, word joiners, byte order marks, soft hyphens, and unusual spacing characters. These can appear through copy-and-paste, document conversion, web editors, or deliberate insertion.
The cleaner performs deterministic transformations in your browser. It does not paraphrase, call a language model, or claim to erase a statistical watermark. That narrow scope is intentional: the result is inspectable and reproducible.
Be careful with multilingual text
Joiners and direction marks can be meaningful in Arabic, Persian, Hebrew, Indic scripts, and emoji sequences. Review cleaned multilingual text before publishing it. The audit lists exactly what it found so you can make an informed choice.
Practical limits
- No absence-of-watermark result proves that a human wrote the text.
- No generic cleaner can verify every provider’s private watermark.
- Rewriting text can change statistical properties, but that is not the same as deterministic removal.
- Provenance systems are one signal and should not be used alone for high-stakes decisions.
Primary sources and further reading
- Google DeepMind: Watermarking AI-generated text and video with SynthID
- Google DeepMind: SynthID Text reference implementation
- Nature: Scalable watermarking for identifying large language model outputs
- Google AI for Developers: SynthID Text guide
Ready to inspect a document? Open the free invisible character cleaner.