Google just gave everyone a way to answer a question that keeps getting harder: is this real, or is it AI? On Wednesday, October 7, the company opened its SynthID detector to the public — a free site where anyone can upload an image, video or audio clip and check whether it carries the invisible watermark Google’s generative models stamp on their output. Google says its existing verification features, already built into the Gemini app and Chrome, now handle more than 1 million checks a day.
The detector hunts for a signal you can’t see. Google’s image models Nano Banana and Veo, its music model Lyria, and tools like Gemini, Flow, ProducerAI and Vids all embed an imperceptible SynthID watermark at generation time. Since introducing the technology in 2023, Google says it has watermarked more than 180 billion images and videos, along with 240,000 years of audio.
This is the most powerful “is this AI?” button anyone has shipped so far.
Until now, the site belonged to specialists. Google tested an early version with journalists, media professionals and researchers starting in May 2025, and the public version is now live globally in English. The tradeoff: you have to sign in, and each account gets roughly 10 checks per day across all supported formats, according to Ars Technica. Google told Ars the cap is deliberate: let people probe the detector without limits, and someone eventually reverse-engineers how to strip the watermark.
It isn’t only Google’s own output under the microscope. OpenAI, NVIDIA and Kakao have adopted SynthID, so the detector can also flag supported media from those companies’ tools. Google announced its partnership with OpenAI in May 2026; OpenAI added SynthID to audio in July and maintains its own verification site. Apple, TechCrunch reports, is expected to add support soon. Microsoft and Meta, meanwhile, run their own watermarking standards — and TechCrunch notes those tools have at times failed to flag content made by their own models.
How the Google SynthID detector works
The mechanics are simple by design. You upload a file; the site accepts a long list of formats, including JPG, PNG, WEBP, HEIC and GIF images; MP4, MOV and WEBM video; and WAV, MP3, OGG, FLAC, AAC and M4A audio — and SynthID scans for the watermark baked into the pixels or riding inside the audio waveform. A verdict comes back quickly, and Google’s sample results show the kind of answer you get: the media was edited with Google AI, for instance, when the watermark traces to one of its tools.
The scale behind the launch is already enormous. In a May 19, 2026 post, Google said SynthID verification inside the Gemini app had been used 50 million times globally, with the capability expanding to Search that day and to Chrome in the following weeks. At that point the company had already watermarked more than 100 billion images and videos and 60,000 years of audio — numbers that have nearly doubled since, according to this week’s announcement, which DeepMind’s Pushmeet Kohli, vice president of science and strategic initiatives, laid out on Google’s Keyword blog.
What the detector won’t tell you
The site is blunt about its limits. “This is not a general AI detector,” its FAQ says — it can only identify media carrying SynthID, from companies that adopted the standard. Upload a clip from a generator that doesn’t watermark, and you’ll get silence, not an answer. Heavy editing, cropping and re-compression can degrade the watermark too.
None of that is a scandal. Watermarking was never sold as a perfect lie detector, and the adversarial math is obvious: any verification system that anyone can query without limits becomes a free training set for defeating it. That is the logic behind the daily quota, and it is also why Google describes the detector as one verification aid among several, not a verdict on everything you see online.
Why this matters
Detection is getting infrastructure instead of press releases. The deepfake problem was never a shortage of research — it was a shortage of buttons ordinary people can press. Google’s coalition is the largest attempt yet at a cross-company answer: watermarks only work at scale if rival labs adopt the same one, and SynthID now has OpenAI, NVIDIA and Kakao on board, with Apple reportedly on the way.
That’s also where the optimism gets qualified. Microsoft and Meta staying on their own standards keeps the ecosystem split, and their tools’ track record on their own models’ output, as reported this week, is not encouraging. The 10-check daily cap is the tell: Google knows the watermark war is adversarial, and this launch is the opening move, not the endgame. If Apple joins and the coalition holds, though, a checkable origin trail becomes the default for most of the AI media people actually encounter — and the unmarked stuff starts looking suspicious by default.
Frequently asked questions
How do I use Google’s SynthID detector?
Go to synthid.com, sign in with a Google account, and upload an image, video or audio file. Each account gets roughly 10 checks per day across all supported formats.
Does Google’s SynthID detector catch all AI-generated content?
No. It can only detect the SynthID watermark, which appears in media from Google, OpenAI, NVIDIA, Kakao and their supported tools. Content from companies using other standards, or no watermark at all, will not be flagged.
Is Google’s AI detector free to use?
Yes, the public site is free. You need a Google account to sign in, and usage is capped at about 10 checks per day to make it harder to reverse-engineer the watermark.
Can the SynthID watermark be removed or fooled?
Google designed the watermark to survive common edits, but heavy cropping, recompression and deliberate tampering can degrade it. Google describes the detector as a verification aid, not a general-purpose AI detector.
Sources: TechCrunch, Unite.AI, Ars Technica, The News (thenews.com.pk)