SynthID and C2PA both label AI content, but in different places. SynthID is Google’s invisible watermark, woven into the pixels of images (and into video, audio and text) made with Google’s AI tools, and it’s designed to survive cropping, filters and compression. C2PA is an open standard for Content Credentials: signed metadata attached to a file that records how it was made and edited, readable by anyone with a verifier, but easy to strip. They’re complementary, and Google now uses both on images from its Nano Banana Pro model.
The short comparison #
| SynthID | C2PA Content Credentials | |
|---|---|---|
| What it is | An invisible watermark in the content | Signed metadata attached to the file |
| Who makes it | Google DeepMind | An open standard from a coalition including Adobe, Google, Meta, Microsoft, OpenAI and others |
| Who adds it | Google’s AI tools | Any tool, camera or platform that implements the standard |
| Media | Images, video, audio, text | Images, video, audio and other files |
| What it tells you | “Made or altered with Google AI” | Which tool made it, what edits followed, whether AI was used, and who signed it |
| Survives screenshots | Designed to | No, a screenshot is a new file |
| Survives crops and compression | Designed to | Often lost on upload or re-save |
| Who can check it | Google’s tools, such as the Gemini app | Anyone, with free verifiers |
| Tamper evidence | Not applicable | Cryptographic signature breaks if content changes |
| Main weakness | Only covers one company’s content; research shows watermarks can be attacked | Easily removed; only as useful as its adoption |
How SynthID works #
SynthID embeds imperceptible signals into content as it’s generated. For images and video, Google says the watermark is designed to stand up to cropping, filters, frame rate changes and lossy compression (Google DeepMind). You can’t see it, and you can’t read it with a metadata viewer.
To check for it, you use Google’s tools. Since November 2025, the Gemini app lets you upload an image and ask whether it was made with Google AI (Google). Google has also been testing a SynthID Detector portal with journalists and media professionals, and says it plans to bring verification to more surfaces, such as Search.
Images made or edited in the Gemini app also carry a visible watermark alongside SynthID (Google).
How C2PA works #
C2PA attaches a manifest to the file. It contains assertions about the content (when and how it was created, what edits were made, whether AI was used), bundled into a claim and signed by the tool or device that made it. Cryptographic hashes tie the manifest to the image, so any change without a C2PA-aware tool breaks the match (C2PA).
Anyone can read it. The Content Authenticity Initiative’s free Verify page shows credentials, Google’s “About this image” can show whether an image with C2PA metadata was made or edited with AI, and Meta and TikTok read credentials to label AI content. Our explainer on what C2PA and Content Credentials are covers the details.
Where each one breaks #
C2PA is easy to lose #
Metadata travels with the file only if every step preserves it. Many platforms strip it on upload, screenshots don’t carry it, and some editors drop it on save. The C2PA FAQ acknowledges that manifests can be separated from files, and describes “soft bindings” (watermarks or fingerprints) as a way to rediscover credentials after removal (C2PA FAQ).
SynthID only covers Google #
A SynthID check can confirm that an image came from Google’s tools. It can’t tell you anything about an image from Midjourney, Stable Diffusion or any other generator. A negative SynthID result isn’t evidence that an image is real.
Watermarks can be attacked #
Academic research on AI image detection has shown that watermarks which change an image only slightly can be removed with a technique called diffusion purification, that stronger watermarks can also be attacked, and that watermark patterns can be added to real images to make them look watermarked (Saberi et al.). That’s not specific to SynthID, but it’s a reason not to treat any watermark as unbreakable.
Why they work better together #
The two systems cover each other’s gaps. C2PA gives a detailed, signed, readable history, but it’s fragile. A watermark is durable but says little. C2PA’s own design allows a watermark to act as a soft binding that points back to a credential stored elsewhere, so a stripped image can have its history recovered.
Google is doing both: images generated by Nano Banana Pro in the Gemini app, Vertex AI and Google Ads carry C2PA metadata as well as SynthID (Google).
What to do when neither is present #
Most images you’ll check carry neither: no credentials because they were stripped or never added, and no SynthID because they didn’t come from Google. That’s when you fall back on a classifier that reads the pixels, plus the usual source and context checks.
Expose AI follows that order on your phone. Its metadata pass reads Content Credentials, XMP and generator tags, and a credible AI claim decides the verdict on its own. When the metadata is missing or says nothing useful, its on-device model judges the pixels and returns Likely real, Uncertain or Likely AI-generated with a confidence meter. It doesn’t read SynthID, so for images you suspect came from Google’s tools, a quick check in the Gemini app is worth adding. See can Google tell if an image is AI.
Frequently asked questions #
Is SynthID part of C2PA? #
No. SynthID is Google’s own watermarking technology. C2PA is an open metadata standard that Google helps develop. Google uses both on some of its images, and C2PA’s design allows watermarks to be used alongside credentials.
Can I see a SynthID watermark? #
No. It’s designed to be imperceptible. You need a detector such as the Gemini app’s verification feature to check for it. Gemini app images also carry a separate visible watermark, but that can be cropped out.
Which is more reliable, SynthID or C2PA? #
Each is reliable within its limits. A valid C2PA credential is strong evidence of origin from any participating tool, but it’s often missing. SynthID is more likely to survive sharing, but only covers Google’s AI content. Neither can tell you about an unlabeled image.
Does Instagram read SynthID or C2PA? #
Meta says it reads the “AI generated” information in C2PA and IPTC metadata to apply its “AI info” label, and it adds its own markers to Meta AI images. See what “AI info” means on Instagram.