Google opens SynthID Detector to everyone, blind spots included

A yes from Google's new detector means something. A no means almost nothing.

Google opened SynthID Detector to the public worldwide on October 7. Anyone can upload an image, video or audio file at synthid.com and check it for a SynthID watermark, according to Google's announcement by Pushmeet Kohli of Google DeepMind. It is in English for now.

It reads more than Google's own marks

The detector reads watermarks from Google and from partners OpenAI, Nvidia and Kakao, Google says, with Apple coming soon. An early version launched in 2025 for journalists and media professionals to test. This one is open to everyone.

SiliconANGLE reports you sign in with a Google, OpenAI or Apple account, and that detection for images from Apple's AI tools arrives in a few weeks.

Google's numbers:

More than 180 billion images and videos carry SynthID to date, plus 240,000 years of audio, since it launched in 2023.

Google says the verification already built into Search, the Gemini app and Chrome handles more than 1 million requests a day. TechCrunch lists the formats the site accepts, including JPG, PNG, WEBP and HEIC images, MP4 and MOV video, and WAV and MP3 audio.

Google DeepMind's SynthID page says its image and video watermarks are designed to survive cropping, filters and lossy compression.

The blind spot is everyone else

The Decoder puts it plainly: the tool only flags content that carries a SynthID watermark. SiliconANGLE adds that Anthropic's Claude, xAI's Grok and many Chinese open-weight models don't use SynthID, and that the tool can't tell content made entirely by AI from content only edited with it.

Think of it as checking a lot number on a pill bottle. If the number is there, you know who made it. If it is missing, you have learned nothing about the pills.

No source we opened gives an accuracy or false positive rate for the public tool.

What to do with it

If you work in a newsroom, in communications or in trust and safety, add it to your first pass. A positive result is strong evidence a file came from a Google or partner model. A negative result should change nothing about how hard you look.

If you publish AI images made with Gemini or ChatGPT tools, assume your audience can now check them in a few clicks. Label them before someone else does.

What to watch

Apple's support is the next date, and it puts the check in front of a much larger set of images. After that, watch whether Meta or Microsoft, which TechCrunch says have their own watermarking standards, agree to let Google read their marks. Every lab that stays out is a lab whose images pass as clean.

Bookmark it. Don't trust a blank.