Google Opens SynthID Detector to Everyone—But ‘Not Detected’ Doesn’t Mean Real

A dramatic video arrives in a group chat. Someone says it is real. Someone else points to a strange shadow and says it is AI. Neither has checked where it came from.

Google has opened another route for investigating that argument. On October 7, it announced global English-language access to SynthID Detector, a tool for checking supported media for an invisible AI watermark. Previously, access had focused on media professionals.

The public launch is useful news. The dangerous assumption would be that the tool can certify everything it fails to flag as authentic.

What Google actually opened to the public

The new announcement covers images, video and audio from Google and participating partners. It names OpenAI, NVIDIA and Kakao, with Apple described as coming soon. Google says its watermarking technology has been applied to more than 180 billion images and videos.

That is a company-reported deployment count. It is not a published accuracy score, and it does not tell us what proportion of the media in your feed the detector can identify.

The practical starting point is synthid.com. We have verified the announcement and technical background, but have not uploaded a test set or independently measured the public service.

A watermark detector answers a specific question

DeepMind explains that SynthID embeds an imperceptible signal when supported AI content is created. Detection looks for that signal. The company designs the watermark to tolerate common changes such as cropping and compression.

This differs from looking at a picture and deciding that the hands are suspicious. The evidence is a signal introduced by a compatible generation system, rather than a viewer’s impression of what synthetic media ought to look like.

Think of the distinction this way: checking for a particular security seal can tell you something when the seal is present. Its absence does not establish the entire history of the object.

Our interpretation is therefore narrow. A result can provide evidence about AI involvement within the detector’s coverage. It does not, by itself, establish the truth of the caption, the identity of the uploader or the circumstances of the scene.

The catch: “not detected” does not mean “real”

Imagine three versions of the same claim: “A bridge collapsed this morning.” One post contains a generated image. Another uses a genuine photograph from five years ago. A third includes an authentic current photograph, but names the wrong city.

Those are different verification problems. A watermark check might help with the first. It cannot, on its own, resolve the date and location claims in the other two. A real camera image can still carry a false story.

There is a second distinction: generated content is not automatically deceptive. An openly labeled illustration for an article is a different editorial choice from an invented photograph presented as eyewitness evidence.

For a newsroom, the useful question is not simply “AI or human?” It is “What is being claimed about this file, and what evidence supports that claim?” The detector can contribute to that investigation without replacing it.

How we would check a viral clip

Here is the workflow we recommend alongside a watermark check. It is an editorial procedure, not a claim that we tested a particular viral video.

1. Write down the exact claim. Separate what the image visibly shows from assertions about who, where and when.
2. Find the earliest available source. A repost with a million views may contain less useful context than the original upload.
3. Preserve the best available file. Record where you obtained it and avoid treating an edited screenshot as equivalent to the original.
4. Run the relevant checks. Use watermark detection as one evidence source, then investigate location, timing and corroboration separately.
5. Describe the conclusion precisely. Explain what was found, what remains unknown, and which claim the evidence actually addresses.

Suppose the tool reports a watermark in a video about an announced product. A careful article would explain that finding. It would not automatically conclude that the product does not exist; the footage might be a disclosed promotional visualization.

The opposite mistake is just as easy. A clean-looking result should not become permission to repost an allegation about a real person.

Why this launch matters beyond the detector

The earlier 2025 introduction described a portal that could identify portions of media likely to contain a watermark. October’s news is the wider access and expanded ecosystem, rather than the invention of watermarking this week.

That change could make evidence checking easier to discuss. Instead of arguing only about uncanny faces, people can ask whether a supported creation signal was found. But the way platforms phrase results will matter: uncertainty needs to remain visible.

For The Bot Post, the editorial implication is simple. Generated covers should be labeled as illustrations, sources should be linked, and uncertain claims should stay uncertain. Readers should not need forensic software to discover how we made an image.

Our existing guide to checking AI-generated images and videos covers the broader problem. This release adds a new public tool to that process; it does not eliminate the need to verify the story attached to a file.

Quick answers

Can SynthID Detector identify every AI image?

It checks for supported watermarks. Treating it as a universal test of every generator would go beyond the mechanism Google describes.

Does an AI watermark mean the whole story is false?

No. It is evidence about media creation. The surrounding factual claims still need their own checks.

Researched October 8, 2026, using Google’s public-access announcement, DeepMind’s SynthID explanation and the original Detector introduction. Deployment figures are Google-reported. Verification examples are hypothetical; we have not benchmarked the detector. AI-assisted, source-checked article.