Why AI images have six fingers — and how to catch them before your audience does
The six-fingered hand is the mascot of AI-generated imagery: instantly recognisable, endlessly screenshotted, and still showing up in paid placements from brands that should know better. Understanding why models get hands wrong explains why the problem persists — and why the fix is process, not prompting.
Models paint appearances, not anatomy
Diffusion models learn what images of hands tend to look like, not what a hand is. A hand is a small object with 27 degrees of freedom that appears in millions of poses, at every scale, frequently self-occluded. From the model’s perspective, "some fingers, roughly parallel, tapering" is what the training data mostly agrees on — the exact count is statistically negotiable.
That is why errors cluster exactly where structure matters and appearance is ambiguous: finger count, joint direction, teeth, ears. The model produces something plausible at a glance because a glance is what it learned from.
Why prompting doesn’t solve it
"Anatomically correct hands, five fingers" in a prompt nudges the distribution; it doesn’t install a skeleton. Newer models are better, but every generation still fails at some rate — and at campaign volume, a 2% failure rate means dozens of shipped hands you didn’t look at closely.
The teams that stopped shipping six-fingered hands didn’t find a magic prompt. They added a check between generation and publication.
A workflow that actually catches them
Zoom is the whole game: anatomy errors survive review because reviewers judge thumbnails. A reliable review inspects every hand, face and joint at 100% scale — which is exactly the sort of diligence humans are bad at sustaining across hundreds of creatives.
This is what Chekr’s anatomy & hands check automates: every person in the creative is inspected limb by limb, each violation gets pinned with a bounding box and severity, and most findings carry a one-click regional fix that leaves the rest of the image untouched. You can try it on one image free — no account needed.
What to do when you find one
Don’t regenerate the whole image for one bad hand — you’ll trade a known defect for a fresh roll of the dice everywhere else. Regenerate the region: inpaint the hand, keep the composition. Then re-check the result, because fixes can introduce their own artifacts; Chekr re-scans corrected images and warns if the integrity score dropped.