Mike HeaversResearch prototype · The findings · Code

Honest AI-image check

Is this image AI?

Drop in an image. We look for proof inside the file, then ask a free AI detector, all on your own computer. You get an honest answer, including "don't know."

Private. Your image never leaves this browser.
One-time download. The detector is 393 MB, saved in your browser after the first check.
Honest. In our test of 363 images, most viral AI fakes could only get "don't know."
Drop an image here or click to choose one · JPEG, PNG, WebP, GIF, AVIF

Checking whether the detector is already saved in this browser…

The image being checked
Checking…

Confirmed means proof found in the image itself: a signed label or a "made by AI" tag. A hidden watermark is also proof, but has to be checked on Google's site, linked below. Likely AI means the detector is at least 95% sure: a guess, not proof. There's no "likely real," because in our test detectors' "real" verdicts on viral images were wrong too often.

Proof in the image

Certain when present, but most shared images carry none, because apps strip it on upload.

Labels in the file

C2PA Content Credentials, the IPTC "made by AI" tag, generator settings
Certain

Reading the file…

Hidden watermark

SynthID, read only by Google's checker
Certain · checked elsewhere

Images from Google, OpenAI, Nvidia and Kakao's AI tools carry an invisible watermark designed to survive compression and re-sharing. In our test, fact-checkers found it in 12 of 59 viral fakes that our detectors missed. Only Google can read it, so this page can't check it for you. Upload the same image to Google's checker; it asks you to sign in.

Found: made or edited with one of those four companies' AI tools. Not found: rules out only those tools, not AI in general.

The detector

An AI model that estimates from the pixels how likely the image is to be AI. A marker in a shaded end means it's at least 95% sure.

See the other detectors (3)· 394 MB more to download

They're saved in your browser too, so this only happens once.

How it works

First it reads the file for signed labels and "made by AI" tags. Then it runs SwinV2, a free detector, in your browser. Its download is saved, so later visits skip it, and checks take a second or two.

What it can't do

Detectors guess, and fakes made to fool people are the ones they miss. It can't read the hidden SynthID watermark itself, since only Google can, but it links to Google's checker. No label doesn't mean real: most apps strip labels on upload.

Where this comes from

A test of 363 images, 123 of them viral and fact-checked. The figures next to each detector are from that test: see the findings. The code is on GitHub. The models are converted from their open-source originals.