AI messes up hands because image generators learn what pictures look like, not how bodies work. Hands take up only a small part of most photos, they appear in an enormous range of poses, and they’re often partly hidden behind objects or other fingers. A generator has no skeleton or finger count to follow, so it reproduces the general look of “hand” and gets the structure wrong. Newer models draw hands far better, but mistakes still show up where hands grip, overlap or sit small in the background.
Reason 1: Generators don’t know anatomy #
A diffusion model builds an image by starting from random noise and refining it toward something that matches the prompt. It learns from vast numbers of images what textures, shapes and colors tend to appear together. What it doesn’t have is an explicit model of a hand: bones, joints, five fingers, a thumb that opposes the rest.
So it produces something that looks like fingers in the right area, with the right skin and shading. Whether there are four, five or six of them isn’t a rule it checks. Counting is something these models were never explicitly taught.
Reason 2: Hands are small in training images #
In a typical photo of a person, the face gets plenty of pixels and the hands get few. Many generators also work on a compressed representation of the image and add detail as they go, so a small region gets relatively little capacity. Faces are well learned because they’re large and usually seen from similar angles. Hands are learned from small, often blurry examples.
Reason 3: Hands have too many poses #
A face seen from the front looks broadly similar across millions of photos. A hand can be open, fisted, pointing, waving, holding a phone, wrapped around a cup, tucked in a pocket or clasped with another hand. Each pose looks completely different in 2D. That variety makes it much harder for a model to learn one consistent structure.
Reason 4: Hands are often hidden #
In real photos, fingers overlap each other, disappear behind objects, and are cut off by the frame. The model sees lots of partial hands and learns that it’s normal for fingers to vanish or merge into things. When it generates a hand holding a glass, it may blend fingers into the glass because that’s close to what it saw.
Reason 5: Captions rarely describe hands #
Text-to-image models learn from images paired with descriptions. Captions say “a woman holding coffee at a café,” not “her right hand has five fingers, thumb wrapped around the handle.” So the text gives the model almost no guidance about hand structure.
What hand errors look like #
| Error | What causes it |
|---|---|
| Extra or missing fingers | No finger count; fingers treated as repeating texture |
| Fused fingers | Overlapping fingers in training images |
| Fingers merging into objects | Grips and occlusion in training images |
| Impossible bends | No joints or skeleton to constrain the pose |
| Two left hands | Left and right look similar in 2D |
| Hands that trail off | Small, low-detail regions at the edge of focus |
Why hands have improved so much #
Generators released since 2023 draw hands much more reliably than early models. Developers trained on more and better images, generated at higher resolution, and used better text encoders and training methods. Some tools also let users regenerate just the hand area until it comes out right. The famous six-fingered hands are much rarer now.
But improvement isn’t elimination. Current generators still slip when:
- Hands interact with objects: holding a pen, playing a guitar, gripping a steering wheel.
- Hands interlock: two people holding hands, fingers laced together.
- Hands are small or in the background, such as in a crowd.
- The prompt asks for unusual poses.
What this means for spotting AI images #
Hands are still worth checking, especially the kinds of interactions above. But a normal-looking hand isn’t evidence that an image is real, because good generators get most hands right. And the person who made an AI image may simply have cropped out, hidden or regenerated the hands.
Treat hands as one clue among several. Look at text, reflections, shadows and geometry too, as covered in AI image artifacts. And remember that real photos have odd hands too: motion blur, wide-angle lens distortion and portrait mode masks can all make a real hand look strange.
When the hands, face and everything else look clean, a detector can check what you can’t see. 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 reads the image as a whole rather than hunting for a single flaw, so it can flag a generated image with perfect hands. For the full checklist, see how to tell if an image is AI-generated.
Other things AI still struggles with #
The same reasons explain several other classic weaknesses:
- Text: strict rules, small details, little guidance in captions.
- Teeth: many small, similar shapes in a row.
- Musical instruments and machines: functional parts that must connect correctly.
- Symmetrical accessories: earrings and glasses that should match on both sides.
The explanation is always the same: a generator learns what things look like, not the rules that make them work.
Frequently asked questions #
Can AI draw hands correctly now? #
Usually, yes. Current generators produce correct hands in simple poses most of the time. Errors are now concentrated in complex interactions, overlapping hands and small background figures, so they’re less reliable as a giveaway than they used to be.
Why does AI give people six fingers? #
The model has no concept of counting. It learned that fingers are repeated similar shapes next to each other and generates “several fingers” without checking the total. Better training has reduced this, but it still happens occasionally.
Is checking the hands enough to spot AI? #
No. Many AI images now have perfect hands, and some real photos have odd-looking hands because of blur or lens distortion. Use hands as one clue alongside other visual checks, source checks and a detector.
Why do AI images hide hands? #
Because hidden hands can’t be wrong. People who generate images often choose or regenerate results where hands are in pockets, behind backs or out of frame. A set of photos where the hands never appear is itself a small clue, as we note in checking a dating match’s photos.