Google DeepMind Opens SynthID Detector to Spot AI Media From 6 Major Providers
Google DeepMind opens its invisible watermark checker to the public, now covering outputs from OpenAI, NVIDIA, Kakao and soon Apple across images, audio and video.
- SynthID Detector is now public at synthid.com, no waitlist required.
- Covers watermarks from Google, OpenAI, NVIDIA, Kakao, ElevenLabs, with Apple coming soon.
- Works on images, video and audio, surviving filters, cropping and compression.
- Google has watermarked 180B+ images/videos and 240,000 years of audio.
- In-product checks via Search, Chrome and Gemini already handle 1M+ requests daily.
- Detection key stays private, programmatic access via Cloud API for select partners.
Google DeepMind has opened the SynthID Detector to the public after limiting access to journalists and researchers through a waitlist. The English-language portal is available globally and accepts images, video, and audio. It searches for invisible watermarks added by supported products from Google and participating providers, including OpenAI, NVIDIA, Kakao, ElevenLabs, and, in a forthcoming integration, Apple.
A signal woven into the file
SynthID embeds a machine-readable signal into generated media at creation time. The watermark lives within image and video pixels or an audio signal, allowing it to remain detectable after many common edits. It does not depend on a visible label or metadata, both of which can disappear when a file is copied, exported, or captured in a screenshot.
Because the signal is distributed through the content, SynthID is designed to withstand operations such as cropping, compression, color adjustment, and audio processing. Robustness varies by format and edit intensity, so detection remains probabilistic rather than guaranteed.
One upload, a localized result
- Sign in to the portal with a supported Google, Apple, or OpenAI account.
- Upload an image, video, or audio file.
- Review the detected watermark and any highlighted regions or segments.
The detector can identify localized signals instead of returning only a file-level verdict. That distinction matters when a generative editing tool changes one object, section, or audio segment while leaving the rest of the source intact.
| Result | Interpretation |
|---|---|
| Watermark detected | A supported generator or editing tool embedded a recognized SynthID signal in all or part of the file. |
| Localized detection | The highlighted region or segment likely contains SynthID-marked material; the result does not characterize the entire file. |
| No watermark detected | The scan found no recognized signal. The file may be unmarked, generated by an unsupported tool, or altered enough to weaken detection. |
The watermark already spans billions of outputs
Google says it has applied SynthID to more than 180 billion images and videos since launching the system in 2023, along with the equivalent of 240,000 years of audio. Verification features in Google Search, the Gemini app, and Chrome now process more than 1 million requests per day, according to the company.
Partners widen the detector’s reach
Watermark detection depends on generators embedding a compatible signal. Expanding SynthID beyond Google products increases the share of commercial AI media that one portal can check.
| Provider | Announced coverage | Status |
|---|---|---|
| Supported outputs from Imagen, Veo, Gemini, and related editing features | Available | |
| OpenAI | Participating image and audio generation products | Available |
| NVIDIA | Participating generative media tools | Available |
| Kakao | Participating generative models | Available |
| ElevenLabs | Participating voice synthesis products | Available |
| Apple | Supported AI-generated media | Forthcoming |
Provider-level support does not imply that every model, product, or historical output from that company carries SynthID. Coverage depends on where each provider has enabled watermarking. Midjourney and many locally run open-source models remain outside the system, leaving their outputs invisible to a SynthID-only check.
Google keeps verification server-side
The public portal exposes detection results while Google retains the verification system and associated keys. Files are uploaded for server-side analysis, and the launch does not provide a public offline verifier or an unrestricted API.
Programmatic access is available through Google Cloud’s AI Content Detection API, which launched with access limited to selected partners. Developers building moderation or provenance workflows therefore need approval for the API; the consumer portal serves manual checks. Teams handling confidential material should review the applicable upload and retention terms before sending files to either service.
C2PA adds a signed record
Supported Google-generated images can also include C2PA Content Credentials, a cryptographically signed manifest that records provenance information and can declare the presence of a SynthID watermark. Applications can read and verify that declaration without access to the private SynthID detector.
C2PA and SynthID address different failure modes. A C2PA manifest provides a signed provenance record but can disappear when metadata is stripped. SynthID embeds its signal within the media, giving it a better chance of surviving routine copying and editing.
The verdict has a narrow meaning
- A positive result identifies a recognized watermark. It indicates that a supported system generated or edited at least part of the file.
- A positive result does not validate the content. A watermark says nothing about whether the depicted event, quotation, or claim is accurate.
- A negative result leaves authenticity unresolved. Unsupported generators, older model outputs, aggressive re-encoding, and adversarial modifications can produce an empty result.
- Detection remains probabilistic. Moderation systems should preserve the score, location data, file history, and other provenance evidence instead of reducing the response to a universal authenticity label.
A public, cross-provider portal gives newsrooms, moderators, researchers, and individual users one place to inspect a growing set of AI-generated media. Its practical value comes from combining SynthID results with C2PA records, source history, reverse search, and conventional verification rather than treating any single detector as conclusive.
Comments
