How to Tell if an Image Is AI-Generated: 9 Signs + a Free AI Image Detector
By Vikram Pagadala
· 7 min read
AI image tools went from obviously fake to frighteningly real in two years. GPT-Image, Gemini's Nano Banana, Midjourney 7 and FLUX can produce a 'phone photo' of a damaged parcel or a spilled meal in seconds. Here is how to tell if an image is AI-generated — by eye and with an AI generated image detector.
9 signs a photo is AI-generated
- 1Text that almost works — labels, receipts, signboards or certificates with warped letters or nonsense words.
- 2Hands, fingers and cutlery that bend, merge or multiply.
- 3Surfaces that are too clean: no dust, scratches, crumbs or sensor noise, even in dim light.
- 4Lighting that is perfect from every side, with soft shadows that don't match the light source.
- 5Repeating or melting patterns in tiles, fabric, grass or rice grains.
- 6Backgrounds that dissolve into smooth blur without real depth.
- 7Odd physics — liquid that doesn't pool, steam in the wrong place, reflections that don't match.
- 8A visible AI watermark, like the small ✦ sparkle Google Gemini adds in the bottom-right corner.
- 9Metadata or Content Credentials (C2PA) saying the image was made with generative AI.
Why your eyes are no longer enough
The newest generators fix most of the classic tells. In blind tests people do little better than a coin toss on good AI photos. That's why AI image detection now relies on models trained on millions of real and generated images, which pick up statistical traces invisible to humans.
How an AI image detector works
Popped! combines several layers so one weak signal can't decide alone:
- Deep-learning AI detectors trained on 2025–26 generators (GPT-Image, Nano Banana, Midjourney, Seedream, FLUX, Grok and more).
- A search for AI tags in EXIF, XMP and PNG data, and a reader for C2PA Content Credentials — ChatGPT and Adobe Firefly images, for example, declare themselves there.
- A Gemini watermark detector that finds the ✦ sparkle even after the photo was forwarded on WhatsApp.
- Frequency and sensor-noise checks that spot the grid patterns and unnatural smoothness of generated images.
What about photos only partly edited with AI?
Tools like Photoshop Generative Fill, Google Magic Editor or Samsung's Galaxy AI change just part of a real photo. These are harder to catch. Popped! looks for Content Credentials that declare AI edits and for patches whose sensor noise doesn't match the rest of the picture, but a small AI edit can still pass — so keep asking for a second photo when money is at stake.
Summary
To tell if an image is AI-generated, look for broken text, hands and physics, check for a watermark or AI metadata, and run it through an AI generated image detector. Use the result as evidence for a decision, not as the decision itself.