For three years, Google has been stamping an invisible mark on the images, video and audio produced by its AI systems. Now it is letting anyone check for that mark. The company’s DeepMind unit opened its SynthID Detector to the public this week, allowing any user to upload a picture, video or audio clip and see whether it carries a watermark from a growing list of AI platforms.
The detector checks content against the watermarking standard Google began applying in 2023. Since then, the company said, it has stamped more than 180 billion images and videos and 240,000 hours of audio. The mark is embedded so that it survives common edits and compression, the company has said, and it is designed to be invisible to the eye and ear.
The tool’s reach is narrower than its name suggests. The detector recognizes content from Google’s own systems as well as from OpenAI, Nvidia and South Korea’s Kakao, with Apple on the list of platforms set to join. The explanation page is explicit that SynthID Detector is not a general AI detector. It only flags material from vendors that have adopted the watermark.
That distinction matters, and Google has been careful to draw it. A tool that claims to detect any AI content invites false confidence and false accusations, because new models and editing tricks can defeat it. By limiting itself to a known watermark, SynthID makes a narrower but more defensible claim: it can say, with some reliability, whether a given piece of media carries a specific marker.
The opening of the tool arrives at a moment when the provenance of digital media has become a public safety question. Election officials, newsrooms and platform moderators have spent years looking for a cheap way to tell whether a video of a politician or a clip of a disaster is real. Watermarks are one answer, though an imperfect one, because they only work when the people who made the content chose to apply them.
Google has been pushing watermarking as part of a wider effort to set standards for AI content. The company has argued that the industry should agree on how to label synthetic media, and that detectors should be available to the public rather than held inside the labs that produce the models. Opening SynthID Detector is a step in that direction, and a way to pressure other vendors to adopt the same standard.
The detector’s limits are worth stating plainly. A watermark can be stripped by a determined adversary, and the tool can return a negative result simply because the content came from a platform that has not signed on. Google has acknowledged both limits and framed the detector as one signal among several, not a verdict.
The broader context is a flood of AI-generated media that has outpaced the tools meant to track it. Every major platform now produces images, video and audio at a scale that would have been hard to imagine a few years ago, and the cost of creating a convincing fake has fallen toward zero. Watermarking is an attempt to keep a record of origin in a medium where origin is no longer obvious.
Google’s bet is that a public detector raises the cost of deception, even if it cannot eliminate it. A journalist who runs a suspicious clip through the tool gets one more piece of evidence. A platform that adopts the watermark gives its users a way to verify content downstream. The system is voluntary, and that is both its strength and its weakness.
Analysts said the move is as much about positioning as it is about safety. By making the detector free and public, Google places itself at the center of the conversation about AI provenance, and sets a bar that rivals may feel pressure to match. Whether the public uses the tool is another matter; the harder problem has always been getting people to check.
The technical approach is an extension of the watermarking method Google built for its image model and has since applied to video and music tools. The mark is embedded during generation by adjusting the model’s output in ways that are imperceptible to a viewer but detectable by software. The detector reverses the process, reading the statistical pattern the embedding leaves behind.
Google is also working inside the broader industry effort to standardize content provenance, alongside the groups that have pushed for labels on synthetic media. The detector fits that push by giving the public a free way to inspect content rather than relying on platforms to police it. The open question is whether watermarking spreads fast enough to matter, since a detector is only useful for the share of content that carries a mark.
For now, the detector is a modest piece of infrastructure with an ambitious purpose. It cannot tell the world which videos are fake. It can tell the world which videos Google and its partners have chosen to mark, and that narrow fact, repeated across billions of files, is the foundation the company hopes to build on.
