SynthID

SynthID is an invisible watermark that Google DeepMind builds into content made with its AI, and the SynthID Detector at synthid.com is where you check for it. Is it free? Yes, but you have to sign in with a Google, Apple or ChatGPT account before it will look at anything. It checks images, video and audio, and it only recognizes AI from Google and its partners: OpenAI, NVIDIA and Kakao, with Apple announced as coming soon.

I tested it with two images, a real photo and one that Gemini made, to see whether it could tell them apart. It did, and the way it did is worth knowing before you trust it.

What SynthID actually is

Visible logos can be cropped out, and metadata gets stripped, for example when someone takes a screenshot. SynthID goes into the content itself: into the pixels of images and video, and into the sound of audio clips. You can't see or hear it, but a detector can. Google also uses it for text in the Gemini app and on the Gemini website, where it works by nudging the words the model picks.

The scale is huge. On October 7, 2026, Google said that since 2023 it has watermarked over 180 billion images and videos, plus 240,000 years of audio. That same day the Detector, which had been an early version for journalists and other media professionals, opened to everyone, worldwide and in English.

How I tested the SynthID Detector

When you open synthid.com, the first thing you get is a terms dialog. Besides Google's usual terms, it lists things you agree not to do, like breaking or bypassing the watermark detection, automating or scraping the tool, and creating several accounts to get around the quotas.

Then I tried to upload a photo without an account, and it stopped me with "Please sign in before detection." You can choose Google, Apple or ChatGPT. Under the upload box it lists whose AI it detects: NVIDIA, OpenAI, Google and Kakao.

I signed in with Google and started with the control: a real photo of the Patio de las Doncellas in the Real Alcázar of Seville, taken by a photographer and shared on Wikimedia Commons. The answer: "SynthID was not detected. It is unlikely that this media was made with AI from a SynthID partner." It added that the image could still have come from a company outside the program, or from before a partner started using SynthID.

Now the hard one. About a week earlier, Gemini had generated an Andalusian patio in another test of mine, but I hadn't kept the original file. All I had was a screenshot of the conversation. So I cropped the image out of that screenshot: a smaller copy that had already been re-saved a couple of times in different formats, which is about as rough as you can be on a watermark. The Detector still said "SynthID was detected. This media was made or edited with Google AI" (it's the screenshot at the top of this review).

Two things I expected and didn't see: it didn't highlight which part of the image carried the watermark, and it didn't show a counter showing how many checks I had left.

What "detected" and "not detected" really mean

"Detected" is the strong answer. It means there's a SynthID watermark, so the content was likely made or edited with an AI model that uses SynthID. Google's FAQ says that in very rare cases it can trigger on content without a watermark, so it's a strong signal, not absolute proof.

"Not detected" is the weak answer. According to the FAQ, it doesn't totally rule out AI. The content may come from a model without SynthID, from before a partner adopted it, or it may have been edited heavily enough to wear the signal down. Google also admits that someone who tries many transformations may find one that doesn't get detected. So use "not detected" as one clue, not as a verdict.

What the Detector can't check

• Text. The new Detector handles images, video and audio only. The early version checked text too; this one doesn't.
• AI from outside the program. Google, OpenAI, NVIDIA and Kakao today, Apple later. Everything else is invisible to it.
• Old content. Anything made before a partner started watermarking won't carry the mark.

Other ways to check for SynthID

You don't have to go to synthid.com every time. Google says the Gemini app, Search and Chrome can check for a SynthID watermark too: you upload the image, video or audio and ask whether it was created or altered by Google AI. These checks already handle over 1 million requests a day.

Gemini · Google

Free with limits · from $4.99/mo

Gemini review →

If you already use Gemini, that's the quickest way to check a picture someone sent you.

SynthID Text, for developers

Here's the bit that confuses people: SynthID Text isn't a website where you paste an essay. It's open-source code from Google DeepMind (Apache 2.0) that developers add to their own language model so its output carries a watermark, and it works with Hugging Face Transformers. If you're a teacher hoping to check homework, this won't help you.

Your files and the rules you agree to

The Detector's privacy notice says your files are processed in real time and deleted as soon as you get the result. To enforce the daily quota, it keeps a digital signature of each file for 24 hours, then deletes that too. I still wouldn't upload anything confidential, but it's a reasonable deal for a free check.

The terms also rule out attacking or bypassing the watermark detection, so I won't be covering how to remove SynthID. My take: if you want to know where an image came from, this is the easiest free check right now, as long as you remember the one thing "not detected" can't tell you.