Back in 2023, spotting an AI-generated image was still fairly straightforward: hands looked like cacti, teeth melted into each other, and backgrounds descended into surreal nightmares. Three years on, the landscape has changed dramatically. Midjourney V7, DALL-E 3, Stable Diffusion XL, and the latest 2025-2026 models now produce visuals that regularly fool even seasoned imaging professionals.
And yet, telltale signs remain. Some are obvious; others require a trained eye. Here are the 7 signals that, in 2026, still most frequently give away an AI-generated image.
The 7 Signs That Give Away an AI Image
1. Hands and Fingers -- the Historic Achilles' Heel
Image generators have long displayed their greatest weakness in hands: extra fingers, twisted knuckles, impossible joints. Recent models have made enormous progress, but hands remain the first place to look. Pay close attention: a discreet sixth finger, an unnaturally long palm, or fingernails pointing in odd directions are strong giveaways.
2. Text and Lettering -- AI Can't Read
When an image generator tries to include text in a scene -- a sign, a newspaper headline, a bottle label -- the result is often visual gibberish. Letters tangle together, words exist in no known language, handwriting looks like random scribbles. Some newer models handle short character strings better, but stay alert for longer or stylized text.
3. Asymmetric Ears and Jewelry
Ears are anatomically complex, and AI struggles to render them consistently. One ear that appears blurred while the other is sharp, an earring that seems to melt into the skin or vanish halfway down, a hoop with no matching counterpart on the other side -- these are clues worth checking. Jewelry in general is revealing: interrupted chains, pendants that defy the laws of physics.
4. Incoherent Backgrounds
A perfectly realistic human subject standing before a backdrop that makes no sense. Doors that lead nowhere, windows with impossible reflections, perspective that shifts between the foreground and the background. AI models often render their subjects with care, then "fill in" the background with less rigour. Make it a habit to examine the entire image, not just the main subject.
5. Reflections in Eyes or Glasses
A human eye reflects its surroundings consistently: a window, a light source, the entire room. In an AI image, reflections in the eyes or in eyeglass lenses are often generated randomly, bearing no relation to the rest of the scene. Reflections that correspond to no visible light source, or two eyes that reflect entirely different environments, are strong indicators.
6. Overly Smooth Skin Textures -- the "Porcelain" Effect
Human skin is imperfect by nature: pores, slight variations in tone, fine hairs, veins visible at the temples. AI-generated images tend to produce uniformly smooth skin, almost translucent, looking more like a figurine than a real person. A total absence of pores or skin texture, especially in a close-up portrait, should raise suspicion.
7. Perfectly Symmetrical Faces
Human faces are naturally asymmetric: one eyebrow slightly higher than the other, one side of the smile more pronounced, one nostril slightly different. Image generators tend toward a symmetry that is too perfect, almost unsettling. If a face looks "too balanced," that may be a clue. Also observe the hairline: it is often far too regular in AI images.
The New Generation of AI Images: Nearly Undetectable
Let's be honest: in 2026, the latest generation of models has closed many of these gaps. Midjourney V7 handles hands with uncanny precision in most cases. Models from the Flux family and the newest "Nano Banana" generations produce portraits that routinely pass every filter listed above.
These new generations rely on increasingly powerful diffusion architectures, trained on billions of images. Overall coherence -- lighting, shadows, perspective, anatomy -- has become their strong suit. The obvious artefacts have all but vanished from "tier 1" images. What remains is a kind of slightly artificial perfection, an absence of randomness that can, paradoxically, become the telltale sign.
How to Train Yourself to Spot AI Images
Reading guides like this one is a solid first step. But real expertise is built through repeated practice, by testing your judgment against a variety of images and receiving immediate feedback.
That is exactly the principle behind Fake or Real: each round presents you with a series of images -- some AI-generated, others genuine photographs -- with a timer. You have just a few seconds to decide. The game covers every difficulty level, from Easy (obvious artefacts) to Ultra Nano Banana (the most realistic images of 2025-2026). After each image, you see the correct answer and the clues you should have spotted.
Ready to put your eye to the test? Hundreds of images await you, from the most obvious to the nearly undetectable.
Play Fake or Real →The more you play, the more your brain learns to process these clues intuitively. After a few dozen rounds, you will start to sense that an image "doesn't feel right" before you have even identified why -- much like an art expert spots a forgery before articulating the reasons.
Summary
Detecting AI images in 2026 demands more attention than it did in 2023, but it is far from impossible. Focus on hands, text, jewelry, backgrounds, and reflections. Accept, too, that some images are simply undetectable to the naked eye -- and that this is precisely why training matters.