Google opens SynthID Detector to everyone — one million checks a day, and a detector that only knows its own family

Google has opened its SynthID Detector site to the public, letting anyone upload an image, video or audio clip and find out whether it carries a SynthID watermark — the invisible signature Google’s generative models have been embedding since 2023. As TechCrunch reported on 7 October, the tool spent the past year with a small test group of journalists, media professionals and researchers; it is now a normal website anyone can visit, and Google says people already make one million verification requests a day across the places SynthID checking is built in, like the Gemini app and Chrome.

The detector site itself accepts everything from JPGs and GIFs to MP4 video and WAV or MP3 audio. Upload a file and it looks for the embedded watermark; results say whether the media was generated by a participating company’s tools. That catch is printed right on the site: this is not a general AI detector, and it can only detect media from companies that have adopted SynthID technology.

Who is in the club

SynthID watermarks ride along with anything generated by Google’s own media models — Nano Banana for images, Veo for video and Lyria for music — and tools with generation built in, including Gemini, Flow, ProducerAI and Vids. According to TechCrunch, OpenAI, NVIDIA and Kakao also support SynthID, and OpenAI maintains its own checking site for its content. Apple, Google says, is reportedly adding support soon.

The detection portal is refreshingly honest about the limits: watermarks are imperceptible to humans and robust against common transformations like cropping or compression, but the system only works on media from participating companies. Microsoft and Meta have their own watermarking standards for their models, and TechCrunch notes these competing tools are not infallible — they often fail to identify content created by their own makers’ models.

Why a closed-club detector still matters

The useful way to read this launch is as infrastructure quietly standardising rather than as a new gadget. Two years ago every lab treated provenance as a competitive feature; now the biggest generators — Google’s family, OpenAI, with Apple reportedly following — are converging on one watermark and interoperable checking. That is the pattern this site has tracked as invisible watermarks spread across the industry, from OpenAI adding SynthID marks to its own image output to the way cloud AI services now sign their work while open-weight local models generally do not. A shared detector only deepens that split: content from the big hosted generators becomes checkable, while anything generated locally remains unverifiable either way.

The split matters for what happens when AI media causes harm. When New Zealand passed its deepfake bill with unanimous first-reading support and regulators elsewhere raced to cover synthetic media, lawmakers quietly assumed detection would exist to support enforcement. A public SynthID portal gives journalists, courts and platforms a free first-pass check for the first time — but only for content from participating companies. A deepfake made with an open-source model carries no watermark to find. Verification infrastructure and generation are in an arms race, and the open side of the race has no watermark to lose.

It is also worth remembering the failure mode TechCrunch flagged: detectors that miss their own companies’ content. SynthID’s resilience claims are strong for common transformations, but no published benchmark yet shows detection rates for heavily edited or re-encoded media at scale. Treat any “the detector says it is real” result as one signal, not a verdict — the same caution we applied to Anthropic’s invisible EU AI Act text watermark, where the interesting questions were about what the mark does and does not prove.

❓ FAQ

What is SynthID Detector? It is Google’s public website for checking whether an image, video or audio file contains a SynthID watermark — the embedded, imperceptible signature added by participating generative AI tools.

Which companies’ media can it detect? Google’s Nano Banana, Veo, Lyria, Gemini, Flow, ProducerAI and Vids all embed SynthID, and OpenAI, NVIDIA and Kakao also support it. Apple is reportedly adding support. Microsoft and Meta use their own watermarking standards, which SynthID cannot read.

Is it a general AI detector? No. Per the site itself, it only detects media generated by participating companies. Content from open-weight or unwatermarked generators cannot be checked.

How do I use it? Upload the highest-quality version of the file at synthid.com. Google says watermarks survive common transformations like compression and cropping, but quality affects detection accuracy.

🔍 THE BOTTOM LINE

A verification tool going from press-room demo to public URL in a year is fast by the standards of AI safety infrastructure, and the million-daily-checks figure suggests the demand was already there. But the honest framing is the one Google prints on its own site: this detects one alliance’s watermarks, not AI itself. The gap between watermarked and unwatermarked generators is now also a gap between checkable and uncheckable media — and that gap, not the website, is the story to watch.