Gemini SynthID text watermark: how detection works
Updated August 20, 2026 · 8 minute read
Google DeepMind's SynthID is the clearest public example of why “AI watermark” can mean more than hidden Unicode. SynthID supports multiple media types, including text. For text, it changes token-selection probabilities during generation and uses a compatible detector to estimate whether the expected pattern is present.
This is not the same as inserting a zero-width character or metadata tag. Scan AI Text can inspect those visible-to-code artifacts, but it cannot promise universal SynthID detection or removal.
How SynthID Text works
Google DeepMind's SynthID overview describes watermarking for AI-generated images, audio, text, and video. Its technical explanation for text and video says the text method adjusts model token probabilities during generation.
The detector analyzes a sequence rather than looking for one special character. Longer and more varied responses generally provide more signal. Short, factual, translated, or heavily edited text can be harder to evaluate. Google describes SynthID as one part of a broader safety approach, not a complete answer to misinformation.
Why detection is configuration-specific
Google's open-source SynthID Text documentation includes a logits processor for watermarking and a Bayesian detector. It explains that the detector must use matching configuration values. Results are probabilistic and use outcome categories rather than absolute authorship proof.
That creates three important limits:
- A generic scanner may not know the configuration used for a particular model or response.
- Editing or translation can weaken a statistical pattern.
- A detector result concerns watermark evidence, not the complete history of who drafted or edited the text.
What Scan AI Text can do with Gemini text
Copying Gemini output into a document can introduce ordinary text artifacts. The local scanner checks hidden Unicode, unusual spaces, direction controls, and supported HTML or Markdown metadata. Each finding includes an exact codepoint or field.
That workflow is useful for clean publishing, copy consistency, accessibility review, and debugging. It is not labeled “SynthID removed” because no matching SynthID detector runs in this app.
Safe review process
- Preserve the original response and source context.
- [Scan the copied text](/) for inspectable artifacts.
- Review every proposed normalization or removal.
- Compare the clean copy with the original.
- Keep provenance or disclosure that your policy or law requires.
If you specifically need SynthID verification, use Google's current supported tools and documentation. Do not infer absence from a clean Unicode report.
Frequently asked questions
Does a zero-width-space result mean Gemini added a watermark?
No. It proves only that the character exists at the reported offset. Many editors, websites, scripts, and copy paths can introduce invisible characters.
Can paraphrasing remove SynthID?
Editing can affect statistical signals, but Scan AI Text does not provide evasion guidance or promise detector bypass. Its cleaning workflow targets inspectable artifacts while preserving meaning.
Is SynthID an AI authorship detector?
It is a watermarking and detection system for compatible generated content. A result should not be stretched into a claim about every contributor, edit, or source.