Real or AI? Google opens its SynthID Detector to everyone, free of charge

What's going on

Google has opened the SynthID Detector to everyone. On this website, you upload an image, a video or an audio file and find out whether it contains an invisible AI watermark. The rollout was announced on October 7, 2026; according to the Japanese platform note.com, Google updated the announcement again on October 29. The site works worldwide but is only available in English for now. It's free. To sign in, you need a Google, Apple or OpenAI account.

Google first unveiled the Detector back in May 2025. At the time, only selected journalists, researchers and other media professionals were allowed to use it. The underlying technology, SynthID, has been around since 2023. It was developed by DeepMind, Google's AI division.

How the watermark works

The key point first: SynthID doesn't look at an image and guess whether it "looks" AI-generated. Other AI detectors work that way, with unreliable results. SynthID instead looks for a digital watermark. That's an identifier a compatible AI tool embeds in the file when it creates or edits it. You can't see it in images and videos, and you can't hear it in audio. Think of it as a signature in invisible ink: nothing shows to the naked eye, and only the right scanner reveals it.

According to Google, the marks are designed to survive common modifications. For images, the company cites cropping and compression, meaning lossy saving to shrink file size, which social networks do constantly. Audio watermarks are supposed to remain detectable even after added noise, MP3 compression or a change in playback speed. According to eWeek, even the 2025 version of the Detector highlighted the image regions and audio segments most likely to contain a watermark. Images from Google's image generator Nano Banana 2 carry SynthID alongside other provenance information. The technology is also built into the Veo video generator and the Lyria music model.

After you upload a file, the site shows a result. Where possible, it also names the AI tool used to create the content. According to TechRepublic, it supports 11 common image formats, three video formats and six audio formats, including JPG, PNG, WEBP, GIF, MP4, MOV, MP3, WAV and FLAC.

Not just Google: partners and numbers

The Detector doesn't only recognize content from Google's own models. It also finds SynthID watermarks from partner companies. Those named are OpenAI, Nvidia and the South Korean company Kakao. Apple is set to join "soon," according to Google. eWeek points out, however, that not every product or output from these partners is automatically covered.

According to Pushmeet Kohli, Vice President of Science and Strategic Initiatives at Google DeepMind, more than 180 billion images and videos have been marked with SynthID since 2023. On top of that comes 240,000 years of audio. The note.com author adds some context: the figure only tells you how much AI content has been marked. It doesn't mean 180 billion fakes have been caught.

The checking itself isn't entirely new, either. You could already query SynthID through the Gemini app, Google Search and the Chrome browser. According to Google, those channels together handle more than a million check requests per day. TechRepublic concludes that the new website is less a new technology than a central hub for an existing system. TIME magazine named SynthID one of the best inventions of the year in October 2026.

What you can use it for: a real-world example

The main Turkish source lists typical everyday cases. A photo in the family WhatsApp group supposedly shows a child in distress. Images claim to show flooding in your own town. A video appears to show a celebrity recommending an investment. Imitated voices and faces of famous people in particular help misinformation spread fast.

That the check can help when it matters is shown, per eWeek, by a case from July: fact-checkers debunked a fake hospital photo of US Senator Mitch McConnell because they found the SynthID watermark in it.

There is a limit, though: according to TechRepublic, you get about 10 checks per person per day. The idea is to prevent anyone from testing similar files over and over to find ways of removing the watermark.

The big caveat: no match doesn't mean "real"

Every source stresses the same point. If the Detector finds no watermark, that doesn't make the content real. There are several reasons:

• Many AI providers use their own watermarks or none at all.
• Open-weight models, meaning AI models anyone can download and run on their own computer, aren't part of Google's system.
• Some edits can make a watermark harder to detect or remove it entirely.

A match has its limits, too. According to TechRepublic, the Detector can't tell whether AI generated an entire file or only edited part of it. So a real photo in which just one detail was altered with AI can also trigger a hit. And the result says nothing about whether the story being told is true. The tool doesn't detect scams, and it doesn't verify who sent a message. The main source therefore advises always cross-checking suspicious content against other reliable sources. The note.com author sums it up like this:

What matters is not to translate the result into a simple choice between "real" and "fake."

She also recommends reviewing the terms of use and data handling before uploading images that contain personal data or third-party rights.

Text is left out

eWeek spotted another detail. The public website only checks images, video and audio, not text. There is SynthID Text, which embeds a watermark into AI-generated text as it's written. But whoever deploys it decides who gets to run the check. There are three levels: fully private (internal only), semi-private (access via a controlled application programming interface, or API) or public (the detector is released for download). So a document can carry a watermark without the recipient being able to verify it. On top of that, text detection is only a probabilistic statement. Heavy rewriting or translation makes the result less reliable. eWeek therefore advises companies to decide in advance whether outsiders may check their documents, and to keep original files and edit logs.

Our take

For everyday users, the rollout is a real step forward in convenience. Instead of hunting for mangled fingers or odd backgrounds, you can simply upload a suspicious file. Those tells are disappearing as generators improve anyway. For newsrooms, fact-checkers and businesses, it's a quick extra verification step.

The fundamental problem remains, though. SynthID only detects what was marked with SynthID. Anyone deliberately creating fakes can use tools without watermarks. With open-weight models, that's easy. A negative result is therefore worth little, and there's a risk: some people may treat a file with no match as "checked and real." Add to that several competing watermark standards. No single tool can check everything online. On the plus side, major players like OpenAI, Nvidia and soon Apple are on board. The more companies use the same system, the more meaningful a match becomes. TechRepublic nails it: the result is a clue, not a final verdict.