Detecting AI Generated Images - Quantable Analytics

You are here: Home › ai › Detecting AI Generated Images

June 9, 2026

Jason Packer

Leave a comment

We’ve just launched a new browser extension called "SlopGuard" which does AI image generation detection locally inside your browser! You can get it from the Chrome Web Store here.

Note: before you click the "Add to Chrome" button, you should know that this is a very large extension (400M), so it will take a few minutes to download and install. This can only be run on desktop.

As the quality of AI generated images becomes nearly indistinguishable from reality, separating the real from unreal becomes increasingly important.

With our new extension, we wanted something that:

Is reasonably accurate. There’s always going to be misses and false positives, but an AI detection model that is wrong much of the time simply isn’t helpful. We’ve been working on this extension for a couple months now and have tried many different models and methods, the majority of which just not reliable. We think we’ve found a combination of things that reaches the important bar of "useful".

Is fast, free, and private . Our goal here is to broadly help users see what’s real and what’s not. The accuracy of the models we use will not be as good as what can be done via a good paid API, but using local models makes it something anyone can use.

Relies on multiple means of identification, and is model agnostic . Since image generation models change so fast, the detection models need to be able to be swapped out as well. Open weight models that are a couple years old are simply not able to detect images created by current generation models.

How does it work?

From any webpage, click the six-fingered hand icon, and it will scan all the images it sees currently on the page, excluding those that are very small. This should take about 1-2 seconds per image on average, and it will either place an "AI" label on the image or add a green border if no AI signals are detected. A grey border around the image means detection failed.

So in the example above it detected the first image as definitely AI, and the second as not. You can also right-click on any individual image to check it that image alone. In cases where’s the infinite scroll (like the Reddit screenshot above), you’ll have to click the hand again to get it scan again once there are new images loaded.

If the image is marked "AI", that means we found metadata. Without definitive metadata the most confidence it will declare is "Probably AI". You can see the outputs of the different layers of detection by turning on the Debug Mode in options.

SlopGuard with debug turned on.

SlopGuard detects AI images in 3 ways:

Level 1: Image Metadata

Metadata is the way things should ideally work! Whether it’s an C2PA signed manifest claim (the gold standard for validating content), or a traditional image metadata standard like EXIF/IPTC, having metadata is really the only way to be 100% sure something is AI. It’s also fast, easy, and open.

Unfortunately metadata tends to get stripped from images online. Sometimes this is purposeful to hide how an image was generated, but more often it’s just how a website works. For example, LinkedIn supports displaying C2PA content credentials on an image, but they strip it from the actual image file they serve. So when you download that image, it’s already lost that metadata. For performance and privacy reasons, stripping metadata from images on the web has been a best practice for many years. Except now we need some of that data back!

Most of the biggest AI image generation tools  (ChatGPT, CoPilot, ElevenLabs, Gemini) are setting metadata of some sort. Grok still does nothing.

Testing with command-line exiftool and c2patool. CoPilot does have a C2PA sig, but it didn’t validate.

Level 2: SynthID Watermarks

Google DeepMind has developed an excellent invisible watermarking system called SynthID that they place on all Gemini-created images.

In May, Google announced other AI image generation tools (including OpenAI and ElevenLabs) are going to be using SynthID to watermark their content as well. ChatGPT has already launched this in images it generates! Although in my testing it appears that currently ChatGPT and Google’s implementations of SynthID are incompatible. In other words, I was unable to detect a Gemini-created SynthID in OpenAI’s image validator and similarly Gemini’s "@Verify AI" couldn’t detect one created by ChatGPT. Currently our extension only detects the Gemini version of Google SynthID, not OpenAI’s variant.

The downside of this method currently from a user perspective is that you have to login to Gemini and upload your image to check if an individual image is AI-generated.

SlopGuard can check lots of images at once with its local model approach. Instead of uploading an image one at a time to Gemini, saying "@Verify AI", and then waiting 20 seconds for a full answer.

While Google has...