SynthID Detector
Watermark detection for AI media, now open to everyone worldwide — with one failure mode you have to design around.
A free tool that checks whether an image, video or audio file carries a SynthID watermark — an invisible marker that Google and its partners embed in AI-generated media. You upload a file and it tells you whether it can find the mark.
What you build with it
- A provenance check inside a moderation pipeline
- Verifying user-submitted media before you publish it
- Journalism and research workflows that need a second signal
- Spot-checking whether an asset came out of a generator
What it gets you
- Covers images, video and audio, not just stills
- Free and open to anyone worldwide
- Partners include OpenAI, NVIDIA and Kakao
- More than 180 billion items watermarked, so coverage is real
- Availability
- Worldwide, English
- Portal
- synthid.com
- Media
- Image · video · audio
- Partners
- OpenAI · NVIDIA · Kakao
- Watermarked
- 180B+ images and videos
Straight from the source
Why it matters
Content provenance is turning into infrastructure, and this is the largest deployment of it. But the useful thing to understand is not what the detector finds — it is what a negative result does and does not tell you, because that is where every trust-and-safety pipeline built on it will go wrong.
A negative result proves almost nothing
The detector finds SynthID watermarks. It does not detect "AI-generated content" — it detects one family of marks, applied by participating tools.
So content from a generator that does not use SynthID returns nothing. Content whose watermark was destroyed by re-encoding, cropping or screenshotting returns nothing. And genuine human photography returns nothing, same as all of the above.
That makes it one-directional evidence. A positive meaningfully raises your confidence that something was machine-generated. A negative is not evidence of human origin, and any pipeline that treats it as a clean boolean is quietly broken.
Probabilistic, and sensitive to how you submit
Google describes the detection as probabilistic rather than a definitive verdict, which is the honest framing for a statistical watermark.
Independent testing by Digital Information World found results differed depending on how the media was submitted — meaning upload format and processing can change the answer for the same underlying file.
If you wire this into a moderation flow, test it against your own pipeline, with your own transcoding, before trusting the output.
Scale, partners and access
Google says it has watermarked more than 180 billion images and videos and 240,000 years of audio since 2023, which is what makes the detector worth anything at all — a watermark scheme is only as useful as its adoption.
Named partners so far are OpenAI, NVIDIA and Kakao, with support for media from Apple's AI tools planned but not yet live.
Access reporting is inconsistent: sign-in is required (Google, OpenAI or Apple accounts), with a daily quota of roughly ten checks reported by one outlet but not confirmed against Google's own announcement.
Reach for it when
- You need a provenance signal inside a moderation or trust pipeline
- Your inbound media mostly comes from the covered partner tools
- A positive detection is useful to you even when a negative is inconclusive
Watch out for
- A negative result is never evidence of human origin — do not treat it as a boolean
- Watermarks do not survive every transform; re-encoding can erase them
- Quota and sign-in details are reported inconsistently; verify against Google directly