How to Check If an Image Was Made by AI
Spotting AI images used to mean counting fingers. That's not the main technique anymore, and the tools that replaced it now live in your browser. Here's what actually works in August 2026, roughly in the order you should try it.
Start in Chrome, it takes two seconds
Since Google's I/O announcement in May, Chrome and Google Search flag AI content using both SynthID watermarks and C2PA Content Credentials. Right-click an image in Chrome and check it. On Android, Circle to Search does the same thing.
Do this first. It's instant and it covers both major provenance standards at once.
What SynthID actually is
SynthID is Google's invisible watermark, and the important detail is that it lives in the pixels — not in the file's metadata. That's why it survives cropping, compression and screenshots, which are exactly the things that wipe out normal image data.
Its problem is coverage. SynthID only catches content from tools that signed up for Google's watermark. Older models and plenty of open-source generators won't trigger it at all. So a positive result is meaningful; a clean one barely tells you anything.
Content Credentials are a different check
C2PA works the opposite way. Instead of marking pixels, it embeds a signed provenance record in the file's metadata — which tool made the image, when, and what edits happened afterward.
If an image carries credentials, drop it into contentcredentials.org/verify and read the history. That's the richest result you can get, because it describes how the image was actually made instead of just answering yes or no.
The catch: a screenshot destroys C2PA metadata. Any image that's been screenshotted, re-uploaded, or converted between formats has probably lost its credentials, and their absence proves exactly nothing.
Now look at it properly
Your eyes still earn their place, especially on images carrying no marks in either direction.
The tells that still hold: garbled text inside the image, fused or mangled fingers, and depth of field that no real lens could produce. Quick to check, and they catch a fair share of lazy fakes.
They're not proof, though. Models keep improving, and a clean-looking image isn't evidence a camera was involved.
Why you stack these instead of picking one
Every method fails differently, which is the whole argument for using all of them.
SynthID misses anything from a non-participating tool. C2PA gets wiped by the single most common thing that happens to images online — someone screenshots them. Visual tells catch sloppy output and miss careful output.
So here's the rule worth remembering: a positive result means something, a negative one doesn't. If a check tells you an image is AI-generated, believe it. If everything comes back clean, all you've learned is that it isn't obviously AI-generated. That's a much weaker claim than it feels like.
The law is catching up
There's a legal layer forming under the technical one. The EU AI Act's Article 50 took effect on August 2, 2026, and it requires visible labels on AI-generated content across Europe. Regulation is usually how coverage gaps like SynthID's eventually get closed — voluntary adoption becomes a compliance obligation, and the holdouts stop being holdouts.
Image: cottonbro studio, via Pexels
