AI Image Artifacts: What to Look For
A reference list of what still fails in generated images, sorted by how useful each one actually is. Two things to keep in mind: the reliable tells have moved on since 2023, and none of them work on the best current output.
Tier 1: still reliable
Ears. Rarely a training focus, and as individual as fingerprints. Generated ears often differ from each other in shape or position, and earrings frequently appear on one side only.
Teeth. Merged into a single block, unnaturally uniform, or changing count between similar images. Still one of the weakest spots.
Where objects meet skin. Glasses on the bridge of a nose, a strap over a shoulder, a watch against a wrist, a hand gripping something. Contact points are hard.
Background text. Signs, spines, labels and packaging degrade into letter-shaped forms — except in models built for typography, which is now a real caveat rather than a footnote.
Tier 2: useful in context
Shadow direction. Compare shadows across the frame. Generated scenes often have shadows that disagree with each other or with the highlights.
Catchlights in eyes. The bright reflections should match each other and correspond to something in the room.
Repeating patterns. Tiles, bricks, fabric weave, crowd faces. Real patterns repeat rigidly; generated ones drift in scale or alignment.
Detail falloff. Real optics concentrate detail at the focal point. Generated images often hold the same density everywhere.
Skin texture. Missing pores and fine lines — though heavy retouching produces something similar, so this is a hint rather than a finding.
Tier 3: obsolete
Finger counting. Current models handle hands well. This was the definitive test in 2023 and is now the most out-of-date advice in the field.
Garbled text everywhere. Some models render clean typography — see our Ideogram page for what that changed.
“Too perfect”. Generators deliberately add grain, chromatic aberration and motion blur now. Polish signals nothing.
Symmetrical faces. Modern output includes natural asymmetry.
Why lists like this have a ceiling
Every artifact here is something a model can be trained to avoid, and each generation of models fixes a few more. That is why the tier-3 list keeps growing.
More importantly, the best current photorealistic output has no visible artifacts at all. Studies consistently show people performing near chance on modern generated faces. If you find nothing on this list, that is not evidence the image is real.
The remaining evidence lives where the eye cannot go — frequency, noise, compression history. That is the argument for running an actual analysis rather than squinting harder, and how detection works explains what gets measured.
Key points
- Ears, teeth and contact points are the most reliable visual tells left.
- Shadows, catchlights and repeating patterns are useful supporting signals.
- Finger counting and garbled text are obsolete — models fixed both.
- Deliberate imperfection means polish signals nothing in either direction.
- The best current output has no visible artifacts; absence proves nothing.
Frequently asked questions
Check an image yourself
Reading about it only goes so far. Our AI image detector runs twelve independent methods and shows you every score, not just a verdict.
More from the blog
Try it on your own image
Free account, 3 checks included, no card required.