Build2 publishersIndependently confirmed2 min readPublished
Google opens SynthID Detector to anyone checking images, video or audio for AI watermarks
Google opened its SynthID Detector to everyone, checking files for the watermark it has put on more than 180 billion images and videos. Publishing teams get usable evidence from a hit, but the check sees only generators that apply SynthID, so a clean file proves little.
The Engineer · Build desk
What happened
- SynthID Detector recognises AI output from Google and from partners including OpenAI, NVIDIA and Kakao.
- Google released an early version of the detector last year for media professionals, and the open version is available globally in English.
- Google says it has also watermarked 240,000 years of audio since SynthID launched in 2023.
- SynthID verification built into Search, the Gemini app and Chrome already handles more than 1 million requests a day.
Compiled by The EngineerSomething wrong?How this is made
Why it matters
- capability Marketing, customer-communications and editorial teams get a check they can record for files from covered generators, in place of judging a file by eye or ear.
- constraint Generators outside the SynthID partner group stay invisible to the check, so its coverage grows only as more vendors adopt the watermark.
- exposure A team that treats a clean result as proof of human origin, or a hit as proof the file is unedited, will clear assets the check never vouched for.
Upload a file and the detector looks for one signal: the imperceptible watermark SynthID places in media from Google and the partners that use it [11][5]. It is a membership test against a known set of generators. Apple's place in that set is listed as "soon" [8], a delivery date familiar to anyone who has read a vendor roadmap.
The two possible results carry different weight. A hit is specific. It establishes that the file carries this particular marker [9]. A miss covers two cases the tool cannot tell apart: a file made without AI, and a file from a generator that never applied SynthID. A write-up of the expansion on dev.to says a file without the watermark should not be treated as human-made, authentic or safe to use for that reason [6].
A hit has limits too. According to the same write-up, a positive result does not show that the media was left unedited after creation, that its content is true, or that the business has permission to use it [7].
Those limits decide where the check goes. Assets reach review from contractors, customers, social platforms and internal tools [16]. I think the check belongs at intake, as a recorded step on each asset. The write-up suggests documenting a detected watermark, and asking the provider about any file whose source or rights are unclear before it is published [14]. A detection can then feed the disclosure and editing rules a team already applies to AI media [17]. A clean result sends the file to the same ownership, consent and accuracy checks as any other external asset [14].
The interface sets the cost. The write-up describes a web-based destination where anyone can check files [2], so each check is an upload. Google's announcement does not include detection accuracy figures, robustness data for cropped or re-encoded files, or an API for bulk checks [10]. Campaign images and clips headed for reposting are the cases the write-up singles out as a risk to brand credibility [18]. For those, one manual upload per asset is a cheap step. I would not build an automated publishing gate around the detector until Google publishes error rates.
What to watch
- Apple shipping SynthID support, which would add its generators to the set the detector can recognise.
- Google publishing detection accuracy figures or robustness data for cropped, compressed or re-encoded files.
- An API or batch interface for SynthID Detector that would let review pipelines check assets without a manual upload.