Skip to main content

SynthID: coverage, open-source status, and how to verify it

Provenance Reference • Updated May 29, 2026

SynthID is Google DeepMind's watermark for AI-generated content. This page maps what it covers across text, image, audio, and video, which parts are open source, the products that embed it, and the tools that can actually verify each layer.

SynthID at a glance

SynthID is Google DeepMind's watermarking technology for AI-generated content. It marks text, images, audio, and video with a signal that is imperceptible to people but detectable by a classifier. Only the text watermark is open source (released October 2024 under Apache-2.0 and built into Hugging Face Transformers). The image, audio, and video watermarks are proprietary, and confirming them requires Google's detector. SynthID answers one question, “was this AI-generated?”, and carries no creator, tool, or edit history. For that richer provenance you need C2PA Content Credentials.

SynthID coverage by modality

SynthID is not one watermark but a family of them, one per content type. Each modality has a different open-source status and a different way to verify it.

ModalityExample modelsOpen source?How to verify
TextGeminiYesSynthID-Text, Apache-2.0 (Oct 2024). In Hugging Face Transformers v4.46+.The open-source library detects text watermarked with your own keys. Gemini text is checked through Google's SynthID Detector portal.
ImageImagen, Gemini, ChatGPTNoDetector is proprietary to Google DeepMind.Google SynthID Detector portal or the Gemini app. No local or open-source option.
AudioLyria, ElevenLabsNoDetector is proprietary to Google DeepMind.Google SynthID Detector portal. No local or open-source option.
VideoVeoNoDetector is proprietary to Google DeepMind.Google SynthID Detector portal. No local or open-source option.

What “open source” means here

The open-source release is SynthID-Text: a reference implementation that lets developers add a watermark to their own language model's output and detect it again. Detection depends on the watermarking keys used at generation time, and a Bayesian detector is trained per key. Google's production keys for Gemini are not public, so the library cannot tell you whether a block of text came from Gemini. It verifies your own watermarked text. To check Gemini text, you still need Google's SynthID Detector portal. Any tool that claims to detect SynthID in arbitrary content without Google's detector is making an unsupported claim.

Products that embed SynthID

Google DeepMind reports that more than 100 billion pieces of content have been watermarked with SynthID. These are the products known to embed it today.

Google Gemini & Imagen

+ C2PA

Image, text

Images and text generated in Gemini carry a SynthID watermark. The Gemini app can also detect SynthID in uploaded images, audio, and video.

Google Veo

Video

AI-generated video from Veo is watermarked frame by frame with SynthID.

Google Lyria

Audio

Music and audio generated with Lyria carry an inaudible SynthID watermark.

OpenAI ChatGPT

+ C2PA

Image

Images from ChatGPT embed a SynthID watermark alongside C2PA Content Credentials (announced May 2026).

ElevenLabs

Audio

AI-generated audio is watermarked with SynthID so it stays identifiable after editing or compression (announced May 2026).

How to verify SynthID

There is no local, in-browser SynthID check the way C2PA Viewer reads Content Credentials. These are the tools and primary sources for checking each layer.

SynthID Detector portal

Proprietary

Google's verification portal for images, audio, and video from Google models. Access is rolling out through a waitlist for journalists and researchers.

Gemini app

Proprietary

The SynthID check in the Gemini app verifies images, audio, and video for a watermark. This is the option available to most people today.

google-deepmind/synthid-text

Open source

The open-source reference implementation for text watermarking and detection. Lets developers watermark and detect their own model output. Apache-2.0.

SynthID Text in Hugging Face Transformers

Open source

The production-grade text implementation, available from Transformers v4.46.0 via a watermarking config and a Bayesian detector class.

Scalable watermarking for LLM outputs (Nature)

Paper

The peer-reviewed paper behind SynthID-Text, published in Nature in October 2024.

Check the C2PA layer locally

SynthID and C2PA cover each other's blind spots. SynthID survives screenshots; C2PA carries the full provenance record. C2PA Viewer reads the Content Credentials manifest from any supported image or video entirely in your browser, with no upload. For the full comparison, see C2PA vs SynthID.

Open the Inspector →

Frequently Asked Questions

Is SynthID open source?

Partly. SynthID-Text, the watermark for AI-generated text, was open-sourced in October 2024 under the Apache-2.0 license and is built into Hugging Face Transformers from version 4.46. The image, audio, and video watermarks remain proprietary to Google DeepMind, and their detectors are not publicly available.

Can I detect SynthID in text from Gemini?

Not with the open-source library alone. Detection depends on the watermarking keys used at generation time, and Google's production keys for Gemini are not public. The open-source release lets developers watermark and detect their own model output. To check whether text came from Gemini, you need Google's SynthID Detector portal.

Can C2PA Viewer detect SynthID?

No. C2PA Viewer reads C2PA manifests locally in your browser and can show when a manifest declares an embedded SynthID watermark, but confirming the watermark signal in the pixels or audio requires Google's proprietary detector. Use the SynthID Detector portal or the Gemini app for that.

What is the difference between SynthID and C2PA?

SynthID is an invisible watermark embedded in the content itself, so it survives screenshots and re-encoding, but it only signals that content is AI-generated. C2PA Content Credentials are signed metadata in the file container that record the creator, tool, date, and edit history, but they are stripped by screenshots and most social platforms. The two layers cover each other's weaknesses.

Which AI products use SynthID?

Google embeds SynthID across Gemini, Imagen, Veo, and Lyria. OpenAI embeds it in ChatGPT images alongside C2PA. ElevenLabs uses it for AI-generated audio. Google DeepMind reports more than 100 billion pieces of content have been watermarked with SynthID.

Does SynthID survive screenshots and editing?

The image, audio, and video watermarks are designed to survive common transformations like cropping, resizing, compression, and re-encoding, because the signal lives in the content rather than the file metadata. The text watermark tolerates light edits but degrades as the text is more heavily rewritten or paraphrased.

See also: How to Verify an AI-Generated Image: C2PA vs SynthID and Supported Devices & Services.