Detection guides
How to detect AI-generated images
AI image generators like Midjourney, Stable Diffusion, and DALL·E now produce photographs that fool casual viewers. The good news: most AI images still leak the same handful of clues. This guide walks through the checks you can do with your eyes, the patterns automated detectors look for, and the cases where neither will give you a clean answer.
1. Look at hands, ears, and teeth first
Generators are trained on billions of images but struggle with structures that vary heavily across viewpoints. Hands are the famous case — extra fingers, fused knuckles, missing thumbs. Ears are nearly as bad: real ears have specific cartilage folds (helix, antihelix, tragus) that AI tends to smear. Teeth often have the wrong count or asymmetric edges.
Zoom into a face on a 'too-perfect' Instagram photo. If the ear looks like a soft blob, treat the rest of the image as suspicious until you have a second signal.
2. Check texture continuity on hair and fabric
Real hair has individual strands that follow gravity and the same lighting. AI hair is often rendered as one matte sheet, with strands that change direction unnaturally between regions.
Fabric tells the same story. Knit textures, embroidery, and printed patterns continue across folds in real photos. AI versions break the pattern at the fold or shift scale randomly.
3. Look for impossible reflections and shadows
Sunglasses, mirrors, polished tables, and water are the highest-yield checks. The reflection should match the rest of the scene. AI generators rarely solve this — reflections often show a different room, the wrong head shape, or simply a blur.
Shadows should all point away from the same light source. If one shadow goes left and another goes down, you are looking at a composite or generation.
4. Read text in the image
Text in AI images is improving but still inconsistent. Street signs, T-shirt logos, license plates, and book covers often contain garbled letters, repeating glyphs, or fonts that drift mid-word. Newer models handle short text but rarely full sentences.
5. Run an automated detector for a second opinion
Eyes are good for the obvious cases. For everything else, run the file through an AI image detector that scores synthetic-texture patterns, compression layering, and edge artifacts. Treat the result as one input alongside provenance — where did the image come from, who posted it, and is the source reliable?
Try the tool
AI Image Detector
Run any suspicious image through the detector to get a Truth Score with the specific signals that flagged it.