Now anyone can test whether images and video are AI-made — but the tool has a blind spot
Google has opened its SynthID Detector tool to the general public. The tool at synthid.com flags watermarks from Google, OpenAI, NVIDIA and Kakao — but tests by several outlets show that AI videos without a watermark slip through unimpeded. A negative result, in other words, is not proof that something was made by a human.
What has happened
On October 7, 2026, Google announced that SynthID Detector, the company's web-based tool for identifying AI-generated and AI-edited content, is now available to everyone. Previously, versions of the tool were limited to journalists and other "media people," according to Engadget. Now anyone can visit synthid.com to check whether a file contains a SynthID watermark.
"Today we are expanding SynthID Detector in partnership with @OpenAI, @NVIDIA, Kakao & soon @Apple as part of an industry-wide effort to make AI-generated content more transparent. Available globally in English, our verification portal lets you check whether media was made with AI," Google wrote on X on October 7, 2026, as quoted by IBTimes.
The tool scans images, video and audio and is free to use, but requires signing in with a Google, OpenAI or Apple account, Engadget reports. It is available globally, in English for now.
How it works
SynthID is built on invisible watermarks embedded directly into AI-generated content — in the pixels of an image, in audio signals, or in the frames of a video. The detector reads these marks and can in some cases highlight which parts of the content the watermark was found in, according to an explanation from Google reproduced by MSN.
In practice, it works like this: You sign in, upload a file, and get an answer within seconds. Ben Patterson at PCWorld describes his test: "I uploaded the images and clips into Google's tool and got answers within seconds: 'SynthID was detected,' with the site correctly telling me that one image was made by OpenAI (ChatGPT), while the other was from Google AI (Gemini)."
The detector recognizes watermarks from a range of partners: Google, OpenAI, NVIDIA and Kakao. Apple is announced as a "soon" partner. A Google spokesperson told CNET that there is no specific date for Apple support, but that it is coming soon. The spokesperson also stressed that collaboration with other companies is crucial, because no single player can solve the problem alone.
Interestingly, the tool appears to be partly prepared for Apple content already. Jeff Carlson at CNET uploaded a photo taken with an iPhone 18 Pro and edited with Apple's Reframe tool, and the detector picked up the embedded SynthID tag. An untouched iPhone photo correctly returned no match. But an AI-generated image from Apple Playground was not detected — even though Apple is announced as an upcoming partner.
The big caveat
Several outlets have tested the tool, and the negative findings are just as revealing as the positive ones.
Patterson at PCWorld uploaded openly AI-generated videos from Reddit and elsewhere — including one of a man guzzling soda from an overturned Coca-Cola truck, and another of kittens lining up to drink milk from a row of baby bottles. "Were these videos AI-generated? Yes. Were they watermarked? No," he writes. The tool returned no match on any of them.
Gizmodo tested an image generated with Grok (xAI), which is not a partner. The answer from SynthID.com was precisely worded, and speaks for itself: "SynthID was not detected. It is unlikely that this media was made with AI from a SynthID partner. However, note that this could have been generated by a non-partner organization, or at a time before a partner implemented SynthID."
In other words: The detector can show that a watermark exists, but a clean result says only that no watermark was found. Patterson concludes explicitly that the tool cannot prove that an image or video was definitively made by a human. The sources do not say which generator was behind the watermark-free videos PCWorld tested, or why they lacked watermarks — only that the result was negative.
Privacy and storage
PCMag points out a practical detail: According to the terms of use, Google stores a digital signature of uploaded media for 24 hours, in order to enforce daily usage quotas and "defend the service against automated attacks." The signatures are automatically deleted after 24 hours, according to the company.
PCMag's Jibin Joseph got a ChatGPT-generated kitchen-redesign image correctly flagged in his test — both when he uploaded the original and a screenshot of it. The latter is relevant in practice: In this case, the watermark survived a screen copy.
The scale, and the timing
Google itself highlights large figures, but these should be taken as the company's own claims. According to The Independent via MSN, citing a blog post by Pushmeet Kohlo, vice president of science and strategic initiatives at Google DeepMind, Google has already watermarked more than 180 billion images and videos, plus audio content equivalent to 240,000 years — a figure The Independent renders as "240,000 years' worth of audio content," while PCMag gives "over 240,000 hours." The discrepancy is enormous (a factor of roughly 8,700), and based on the available sources it cannot be resolved. Readers should therefore treat both stated figures with caution.
"This tool comes in addition to our built-in verification features in Search, the Gemini app and Chrome, which now regularly handle over 1 million requests daily," Kohlo writes in the blog post, as quoted by The Independent via MSN.
Google also writes in the blog post that SynthID has been in use since 2023, and the company acknowledges that it is not a universal solution, according to IBTimes.
The timing is no accident. The launch came just days after OpenAI launched an invisible watermark for text generation in ChatGPT, and in the middle of requirements from the EU's AI Act that output from generative AI must be machine-identifiable, The Independent writes. The context, in other words, is both technological and regulatory: Watermarking is moving from something individual players experiment with to something becoming industry practice — and, in the EU's case, a legal requirement.
What this means for the reader
SynthID Detector is a useful tool — but a specialized one. It is free, fast, and appears to reliably pick up watermarks from the partners' tools. For journalists, newsrooms and ordinary users who want to check whether an image comes from, say, ChatGPT or Gemini, it provides a concrete, verifiable answer within seconds.
But it is not a general authenticity test. A video without a watermark from an unknown AI generator, or content from a service that does not use SynthID, will return a negative result — and thus false reassurance. A practical rule of thumb from the first tests: "SynthID detected" is solid evidence. "No SynthID detected" is merely the absence of a watermark, not proof of human-made content.
The open questions are just as concrete: When will Apple support arrive (per CNET, there is no date)? Will more generators embed watermarks — or choose not to? And will watermarks survive compression, cropping and further editing in practice? The answers will determine whether tools like this become a genuine guiding star in the media landscape, or just one of several boxes to tick.