The AI image artifacts worth looking for are garbled or almost-right text, broken geometry (railings, tiles, windows), hands and small objects that merge or multiply, light and shadows that disagree, reflections showing the wrong scene, and skin that looks waxy or airbrushed. They happen because image generators learn what pictures tend to look like, not how the world works. The catch is that the newest models avoid most of these errors most of the time, so finding an artifact is evidence of AI, while finding none isn’t evidence of anything.
Why do AI images have artifacts at all? #
A diffusion model, the kind behind most current generators, starts from random noise and refines it step by step toward an image that matches the prompt. It has no 3D model of the scene, no rulebook for spelling, and no idea that a hand has five fingers. It has learned statistical patterns from a huge number of images, and it reproduces what’s typical.
That works brilliantly for things that vary smoothly and appear often, like faces seen from the front, skies and scenery. It breaks down on things that follow strict rules, such as text, perspective, mechanical parts and the count of fingers, and on small details that take up few pixels in training images.
The visible artifacts, by category #
| Where to look | Typical artifact | Why it happens |
|---|---|---|
| Text and signs | Letters that melt, words that are almost words | Spelling is a strict rule; the model learned letter shapes, not language |
| Hands and fingers | Extra or fused fingers, impossible grips | Hands vary enormously in pose and are often partly hidden in photos |
| Geometry | Railings with uneven spacing, tiles that don’t line up | No 3D scene underneath, only local plausibility |
| Light and shadow | Shadows in two directions, faces lit from nowhere | Lighting is assembled region by region, not from one light source |
| Reflections | Mirrors showing a different room, or nothing | Reflections require copying the scene consistently, which models do poorly |
| Accessories | Mismatched earrings, glasses with one arm | Symmetry across a face isn’t enforced |
| Backgrounds | Crowds of half-formed faces, objects that dissolve | Less attention is spent on small, distant details |
| Skin and surfaces | Waxy, poreless skin; plastic sheen | Models smooth toward an average “good-looking” texture |
| Mechanical detail | Bicycles, keyboards, guitars, clock hands that don’t work | Functional parts have rules models don’t know |
Text #
Text used to be the easiest giveaway. Newer models spell short phrases well, especially when the prompt asks for specific words. The errors have moved to the edges: the small print on a bottle, the second line of a shop sign, the letters on a license plate in the background. Look at text the prompt probably didn’t ask for.
Hands #
Hands have improved a lot, but they’re still worth checking, especially where they touch something: fingers wrapped around a cup, interlaced fingers, a hand resting on another person’s shoulder. We explain the reasons in why AI gets hands wrong.
Geometry and physics #
Follow straight lines across the whole image. A balcony railing that changes spacing halfway, stairs that don’t connect, a table leg that ends in mid-air, a car whose wheels don’t match: these come from a model that checks each region for plausibility without enforcing that the whole scene holds together.
Light #
In a real photo, one sun casts every shadow in the same direction. Check the shadows of people, poles and objects. Check that a face lit brightly from the front makes sense given where the light seems to be. Generated scenes often have a flattering, even glow that no single light source would produce.
Older GAN artifacts vs newer diffusion artifacts #
Many fake profile photos still come from older GAN models, which have their own signature: eyes placed in the same spot in every image, backgrounds of colored mush, a faint second outline around hair, and teeth that blur together.
Diffusion images look different. They’re more coherent, more polished, and more likely to look like a professional photo or a piece of digital art. Their tells are subtler: the too-perfect lighting, the glossy skin, the background that’s blurred so smoothly that nothing in it is identifiable. For illustrations and paintings, where there’s no camera to compare against, see how to tell if art is AI-generated.
The artifacts you can’t see #
Generators also leave traces that people can’t see at all. Early research found that images from one GAN shared statistical artifacts with images from other, unseen GANs, enough that a classifier trained on one could detect many (Wang et al., CVPR 2020). These live in the fine structure of the image: patterns in noise, in how neighboring pixels relate, in frequency content left by the way a model upsamples.
AI image detectors are built to read these invisible traces, which is why they can flag a generated image that looks flawless to a person. How AI image detectors work explains the approach.
That’s also where a detector earns its place in your routine. When you’ve zoomed in and found nothing wrong, Expose AI runs a neural network trained on the visual fingerprints of AI image generators, entirely on your phone, and returns Likely real, Uncertain or Likely AI-generated with a confidence meter. It also reads any Content Credentials or generator tags in the file. Invisible traces weaken when an image is heavily compressed or screenshotted, so the app says Uncertain instead of guessing when the signal is too thin.
Real photos have artifacts too #
Before you call something AI, remember what phone cameras do to real photos:
- Portrait mode blurs backgrounds with a software mask that can cut through hair, glasses and ears, producing edges that look generated.
- Night mode and HDR merge several frames, which can smear moving objects and create halos around bright edges.
- Beauty filters smooth skin to the same poreless finish that generators produce.
- Heavy compression from repeated uploads creates blocky, smeared textures.
A single odd detail in a phone photo is often the camera’s own processing. Look for several independent artifacts before concluding anything, and see why real photos get flagged as AI for the same problem from a detector’s side.
How to use artifacts in practice #
- Open the largest version of the image you can find.
- Check text, hands, geometry, light and reflections, in that order.
- Count independent artifacts. One is a maybe. Three unrelated ones is a strong signal.
- If you find none, don’t conclude it’s real. Move on to source checks, metadata and a detector.
Frequently asked questions #
What are the most common signs of an AI-generated image? #
Garbled background text, inconsistent shadows, geometry that doesn’t line up, hands that merge with objects, and waxy, poreless skin. None is proof alone, and the best current generators often show none of them, so combine visual checks with source checks and a detector.
Do AI images still have six fingers? #
Much less often than in 2022 and 2023. Current models usually draw hands correctly in simple poses, but errors still appear when hands grip objects, interlace, or appear small in the background.
Can AI artifacts be removed? #
Some can. People fix visible errors by regenerating parts of an image or editing them by hand, and compression can hide the fine traces detectors read. That’s one reason no single check, visual or automated, is reliable on its own.
Why does AI make backgrounds blurry? #
Blurred backgrounds are common in professional photos, which make up a lot of training data, and blurring hides details the model would otherwise get wrong. The blur is often unusually smooth, with no identifiable objects, which is itself a clue.