Spotting AI-Generated Images: The Signs That Still Work

You've probably heard the advice: "check the hands." Count the fingers. Look for melted text. Zoom in on the teeth.

A year or two ago, that worked. Today it mostly doesn't. The newer AI generators have fixed most of those slip-ups, and if "zoom in and look for weird hands" is your whole plan, you're already behind.

So what should you look at instead? Let's go through what still works, and be honest about what doesn't.

The old giveaways that have faded

• Hands and teeth. Much better now, though not perfect. Don't rely on them.
• Text inside the image. Improving quickly.
• "It looks too perfect." Plenty of real photos are retouched until they look perfect too.

Looking closely can still catch an obvious fake. But it fails exactly when it matters most, which is when someone is trying to fool you.

Signs that hold up better

1. The photo's hidden details

Photos taken with a phone or camera usually carry hidden information, such as the device model and when the photo was taken. AI images normally have none of that, or they carry traces of the software that made them.

One caution. Many apps, like WhatsApp, strip those details from real photos. So "no details" is a weak clue. It only means something when other clues agree.

2. How the image was saved

A real photo saved once has an even compression pattern all over. A picture that was edited and saved again, or built from several images, may have patches that were compressed differently. A technique called error level analysis can make that visible.

3. The grain

Every camera sensor leaves a faint pattern of noise in its photos. AI generators produce a different kind of pattern. You can't see the difference, but software can measure it.

4. Pasted-in faces

When someone swaps a face into a photo, the edges, the skin texture and the direction of the light often don't match the rest of the picture. This matters a lot for edited ID photos and fake profiles.

5. A trained model's opinion

A model that has studied huge numbers of real and generated images can pick up patterns no person could describe. It's powerful, but it can be fooled, and it can wrongly flag an unusual real photo. So treat it as one voice in the room, not the judge.

6. The story around the picture

Often the strongest clue isn't in the pixels at all. Does this image appear elsewhere online? Does it match what the person told you about it? Does the account that sent it behave normally? A detector helps, but common sense still matters.

Why a list of signals beats a single percentage

Imagine a tool just says "87% AI." What do you do with that? You can't explain it to the person you're about to reject, and you can't tell whether it's right.

Now imagine it says: no camera details, uneven compression in the lower half, an unusual noise pattern. That's something a reviewer can weigh and explain. This is why our image scan lists the signals that fired, each with a severity and a short explanation.

A review routine you can follow

1. Run the scan and read the signals, not just the verdict.
2. Count how many separate signals agree.
3. For anything important, ask for the original file, or for a fresh live capture. For faces, a liveness check works well.
4. Write down what you decided and why.
5. Never penalise someone because of one weak signal.

Try it on your own pictures

Run a free check from the homepage. For the bigger picture, including video and text, read Is this photo real?. And if you'd like to build it into your own product, see the API reference.