SynthID Text

At a glance

Text

Content CredentialsA signed record (C2PA) of how the file was made, sealed so any change shows.

Not stated. None of Google's pages listed under Sources read says so.

Invisible watermarkA signal woven into the pixels, frames, audio or wording, meant to survive copying.

Google says SynthID Text has been open-sourced, with an implementation in Hugging Face Transformers v4.46.0 and later and a reference implementation on GitHub.

SynthID: Tools for watermarking and detecting LLM-generated TextGoogle AI for Developers (Responsible Generative AI Toolkit)

Quoted passage introduction

SynthID Text has been open sourced to make watermarking for text generation available to developers. You can read the paper in Nature for a more complete technical description of the method. A production-grade implementation of SynthID Text is available in the Hugging Face Transformers v4.46.0+ , which you can try out in the official SynthID Text Space . A reference implementation is also available on GitHub that may be useful for open source maintainers and contributors looking to bring this technique to other frameworks.

AI label in metadataA plain metadata field (IPTC Digital Source Type) saying AI made the content. Anyone can add or remove it.

Not stated. None of Google's pages listed under Sources read says so.

Visible markA mark you can see, such as a logo or a sparkle in a corner.

Not stated. None of Google's pages listed under Sources read says so.

Tool named in metadataThe tool or model written into ordinary file fields, such as Software or Creator.

Not stated. None of Google's pages listed under Sources read says so.

Limits and detection

Checking tool from the makerA way the maker offers to check a file for its traces.

Google says SynthID Text detection is probabilistic: its Bayesian detector returns watermarked, not watermarked or uncertain.

SynthID: Tools for watermarking and detecting LLM-generated TextGoogle AI for Developers (Responsible Generative AI Toolkit)

Quoted passage Watermark detection and verifiability

Watermark detection is probabilistic. A Bayesian detector is provided with Hugging Face Transformers and on GitHub . This detector can output three possible detection states—watermarked, not watermarked, or uncertain—and the behavior can be customized by setting two threshold values to achieve a specific false positive and false negative rate.

SynthID: Tools for watermarking and detecting LLM-generated TextGoogle AI for Developers (Responsible Generative AI Toolkit)

Quoted passage introduction

SynthID Text has been open sourced to make watermarking for text generation available to developers. You can read the paper in Nature for a more complete technical description of the method. A production-grade implementation of SynthID Text is available in the Hugging Face Transformers v4.46.0+ , which you can try out in the official SynthID Text Space . A reference implementation is also available on GitHub that may be useful for open source maintainers and contributors looking to bring this technique to other frameworks.

What the maker says about limitsIn the maker's own words.

Google says SynthID Text is less effective on factual responses, loses detector confidence when text is thoroughly rewritten or translated, and is not designed to directly stop motivated adversaries from causing harm. The documentation page was last updated April 9, 2025.

SynthID: Tools for watermarking and detecting LLM-generated TextGoogle AI for Developers (Responsible Generative AI Toolkit)

Quoted passage Limitations

SynthID Text watermarks are robust to some transformations—cropping pieces of text, modifying a few words, or mild paraphrasing—but this method does have limitations. Watermark application is less effective on factual responses, as there is less opportunity to augment generation without decreasing accuracy. Detector confidence scores can be greatly reduced when an AI-generated text is thoroughly rewritten, or translated to another language. SynthID Text is not designed to directly stop motivated adversaries from causing harm.