Can a generic checker detect SynthID Text?
No. SynthID Text is a probabilistic watermark created during generation by adjusting token scores. A matching detector analyzes token sequences with the relevant configuration; there is no invisible character to reveal or delete.
The local checker labels SynthID as not testable. It can still show literal Unicode and C2PA credentials in the same text, but a zero hidden-character result says nothing about the SynthID token pattern.
How SynthID Text creates a watermark
Google describes SynthID Text as a logits processor applied during generation. A pseudorandom g-function changes token scores after top-k and top-p processing. The output remains ordinary text, but the sequence contains a statistical signal that the matching detector can score.
| Component | Role | Why a generic scan lacks it |
|---|---|---|
| Tokenizer | Converts the visible text back into model tokens. | Token boundaries differ between model families. |
| Secret keys | Parameterize pseudorandom watermark scores. | Google recommends storing each configuration securely. |
ngram_len | Balances detectability against robustness to editing. | The detector must match the generation configuration. |
| Bayesian detector | Returns watermarked, not watermarked, or uncertain. | It is probabilistic rather than an exact substring search. |
Open source does not mean one universal detector
Google has open-sourced SynthID Text for developers and documents a production implementation in Hugging Face Transformers. The same documentation explains that watermark configurations use keys and that organizations choose whether their detector remains private, is exposed through an API, or is released publicly. Having the algorithm does not supply every model provider's configuration.
What our 10,000-response test found
We scanned 3,000 watermarked Google examples and 3,000 clean Google controls alongside 4,000 outputs from other models. Both Google groups showed zero supported hidden Unicode markers. That is the expected result for a token-probability watermark, not evidence that SynthID failed. Read the dataset breakdown and methodology.
Can SynthID Text be removed?
There is no attached payload for this tool to strip. Google reports that detector confidence can fall after thorough rewriting or translation, while mild edits may preserve the signal. Those transformations change the words and do not provide a general proof that the watermark is gone. Removing zero-width spaces or metadata is unrelated.
How this checker prevents a false result
The text interface uses separate rows: literal and C2PA signals receive an exact found-or-not-found result, while Claude and SynthID remain “not testable.” The evidence report also records provider-keyed statistical text watermarks under notTested. This prevents “0 characters found” from being converted into “no SynthID.”
Primary sources
- Google AI: SynthID Text application, detector, and limitations
- Google DeepMind: SynthID overview
- Google DeepMind: open-source SynthID Text implementation
- AI text watermark checker coverage
Frequently asked questions
Can I detect Gemini SynthID Text by pasting text into a Unicode checker?
No. SynthID Text is encoded through model token probabilities, not as a hidden Unicode character. Detection requires the compatible tokenizer, configured keys and n-gram settings, and a matching probabilistic detector.
Is the open-source SynthID detector a universal Gemini checker?
No. Google provides technology for model developers to watermark and detect outputs using their own configurations. The documentation says configurations should be stored securely and describes private, API-based, and public detector deployment choices.
Can removing zero-width characters erase SynthID Text?
No. Removing unrelated hidden characters does not change the token sequence that carries the statistical watermark. Thorough rewriting or translation may reduce detector confidence, but it also changes the prose and is not deterministic removal.
Why did watermarked and clean Google samples both show zero hidden markers?
Because the hidden-character scanner examined Unicode code points while SynthID Text lives in token-selection probabilities. In our 10,000-response study, 3,000 watermarked Google samples and 3,000 clean controls both produced zero supported literal markers.