AI Label RemoverWhat Is SynthID? Google's Invisible AI Watermark, and Why Removing Metadata Can't Touch It
SynthID is an invisible watermark that Google DeepMind adds to content made with Google's AI tools — images, video, audio and text. It is built into the content itself rather than attached to the file as metadata, so you can't see it, and you can't strip it out the way you strip EXIF or C2PA data.
That last point is the reason for this article. We make a metadata remover, and the honest answer to "will it remove SynthID?" is no — not ours and not any other metadata tool. Below is what SynthID is, how it differs from metadata, which images carry it, how to check for it, and what that means if your photo is being labelled as AI.
What is SynthID?
Google DeepMind introduced SynthID in August 2023 for images made with its Imagen model. It has since been extended to video (Veo), audio (Lyria), text from Gemini, and images from the Gemini app. The text version was open-sourced in 2024.
SynthID is no longer Google-only. Google lists OpenAI, NVIDIA, Kakao and Apple as partners whose AI-generated content its detector can identify (Apple's support was announced as "coming soon"). In October 2026, Google said it had watermarked more than 180 billion images and videos with SynthID.
How SynthID works
When a supported tool generates an image or a video, a model embeds a pattern into the content at the moment it is created. The change is designed to be imperceptible — the picture looks the same to a person — but a matching detector model can recognise it and report whether the watermark is present, absent or uncertain.
Google says the image and video watermark is designed to survive common changes such as cropping, filters, frame-rate changes and lossy compression. That is the whole point of a watermark: metadata is easy to lose, so the signal is put somewhere ordinary editing doesn't reach.
Text works differently. SynthID for text subtly adjusts which words a model chooses as it writes, and detection looks for that statistical pattern. Everything else in this article is about images.
SynthID vs C2PA metadata
SynthID is often mentioned alongside C2PA metadata (Content Credentials), but they are different layers that do different jobs:
| SynthID watermark | C2PA / Content Credentials | |
|---|---|---|
| Where it lives | In the image content itself | A signed record stored in the file's metadata |
| What it says | "Made or edited with a supported AI tool" | Which tools made and edited the file, and how |
| Who can read it | SynthID detectors (Google's and its partners' checks) | Anyone, with an open C2PA reader |
| Removed by stripping metadata? | No | Yes |
| Designed to survive re-compression? | Yes | No — re-saving or screenshotting often drops it |
Many AI images carry both. Google adds C2PA metadata alongside SynthID to images from its newer image models, and OpenAI pairs Content Credentials with SynthID too. Removing the metadata takes away one layer and leaves the other.
Why removing metadata doesn't remove SynthID
A metadata remover such as AI Label Remover works on the file's containers, not its picture. It deletes or blanks the blocks where C2PA, XMP, EXIF and IPTC data are stored and leaves the compressed image data untouched — in fact, after cleaning, it reads the file back and confirms the image data is byte-for-byte identical to the original.
That guarantee is exactly why SynthID stays. The watermark is part of that image data, so a tool that never changes the image data never changes the watermark. AI Label Remover doesn't detect SynthID either; it only reads metadata.
The same applies to the usual shortcuts. Converting to another format, re-saving at a different quality or taking a screenshot re-encodes the picture, but SynthID is designed to survive exactly that kind of change. Researchers have published attacks that try to weaken image watermarks by altering the pixels — Google disputes the reported results — but that is image manipulation, not metadata removal. It is not something we do or recommend.
Which images carry SynthID?
Content made or edited with Google's AI tools, including:
- The Gemini app — images it generates or edits, and videos made with Veo.
- Google Photos — photos edited with Reimagine in Magic Editor have been marked since February 2025. Google notes that very small edits may be too minor to carry a detectable watermark.
- Google Cloud — images from Imagen on Vertex AI, where SynthID first launched.
Plus content from partners that have adopted SynthID, such as OpenAI and NVIDIA.
An ordinary camera photo doesn't carry SynthID — unless you edited it with one of these tools.
How to check if an image has SynthID
- synthid.com — Google's SynthID Detector, free and open to everyone since October 2026. It checks images, video and audio for the watermark from Google and its partners.
- The Gemini app — upload a file and ask "Was this created with Google AI?" (or type
@synthid). It only recognises content from Google's own AI tools, accepts one file of up to 100 MB at a time, and has a daily limit on checks. - Google Search and Chrome also have built-in SynthID checks.
Read a negative result carefully. "No SynthID found" means the content wasn't made or marked by Google or a partner — not that it wasn't made with AI. Other AI tools don't use SynthID at all.
To see the other half — whether the file carries C2PA, XMP or IPTC metadata — drop it into AI Label Remover. It lists what's in the file, right in your browser, without uploading it.
Can SynthID trigger an "AI info" label?
Possibly, depending on the platform. When Meta introduced its AI labels in 2024, it said it reads industry-standard signals — C2PA and IPTC metadata — and was working on detecting invisible markers from other companies. Platforms don't publish exactly which watermarks they check today, so you shouldn't assume a watermark is ignored.
In practice:
- If a photo was flagged because of metadata an app left behind, cleaning the file removes that signal.
- If the image was made or edited with a SynthID-enabled tool, cleaning the metadata won't change the watermark, and any platform that checks for it can still identify the image.
No tool can guarantee a label won't appear. Platforms also run their own checks and use what you disclose when posting.
What this means for you
A real photo labelled because of an editing app. If you retouched a photo in an app that writes Content Credentials — Photoshop's Generative Fill is a common example — the label most likely comes from metadata. Unless you used a SynthID-enabled tool, there is no watermark involved, and removing the metadata deals with the signal.
A photo edited with Google's AI tools. It may carry SynthID. Removing metadata cleans up the file and your privacy details, but the watermark stays.
An image generated with AI. It is likely to carry SynthID or another watermark, and the label is accurate. Platform rules and laws such as the EU AI Act may require you to disclose AI-generated content. Removing metadata is meant for labels that don't belong — and for privacy, since the same step removes GPS location and device details. Only clean files you own or have permission to edit (see our Terms of Service).
Key takeaways
- SynthID is Google DeepMind's invisible watermark for AI content; OpenAI, NVIDIA, Kakao and Apple are among its partners.
- It lives in the image itself, not in the file's metadata, and is designed to survive cropping, filters and compression.
- Metadata removers — including AI Label Remover — can't remove SynthID, and converting or screenshotting doesn't either.
- Check for it at synthid.com or in the Gemini app; a negative result only rules out Google and partner tools.
- C2PA metadata and SynthID often travel together: cleaning the file removes the first, never the second.