Google Opens Its SynthID AI-Watermark Detector to Everyone

A new public site lets anyone check whether an image, video or audio file carries an invisible SynthID watermark. Google says it has watermarked more than 180 billion images and videos and 240,000 years of audio since 2023 — but the detector only finds its own marks.

Key Takeaways

A new public site lets anyone check whether an image, video or audio file carries an invisible SynthID watermark.
Google says it has watermarked more than 180 billion images and videos and 240,000 years of audio since 2023 — but the detector only finds its own marks.

AI synthesis3 sources0

Google has opened its SynthID Detector to the public, turning an internal verification tool into a site anyone can use to check whether a file was made with AI. The detector looks for the invisible watermarks Google has embedded in output from models such as Gemini, Veo and Lyria since 2023, and accepts JPG, PNG, MP4 and MP3 files. A watermark, Google argues, needs a matching reader to be useful: the company says it has now embedded SynthID in more than 180 billion images and videos and in 240,000 years' worth of audio.

The tool is deliberately narrow. It does not judge whether a file looks AI-generated — it only confirms the presence of a SynthID watermark, meaning content from providers that do not use the system simply comes back unmarked. Rival detection schemes share that limitation, generally verifying only their own watermarks. Google says partners including OpenAI, Nvidia and Kakao already embed SynthID and that Apple is expected to join; for text, OpenAI relies on its own system.

The detector is also being woven into other Google products. SynthID verification is available in Google Search, the Gemini app and Chrome, where the company says it handles roughly one million checks a day. Google frames watermarking as a provenance tool — labelling AI-made media at the source rather than trying to spot it after the fact. That approach has an obvious ceiling: it only works as long as the watermark survives editing, cropping and re-encoding, which is precisely what people distributing misleading media tend to do.