Why AI image generators struggle with hands and text

The two classic tells of an AI image

For a couple of years, two flaws gave away almost any AI-generated picture: a hand with six fingers, and a shop sign covered in confident-looking nonsense. People joked that AI could paint a photorealistic dragon but not count to five or spell a single word.

It is worth understanding why, because the reason is not that the models were lazy or dumb. Both problems come from the same root: a text-to-image model does not understand objects or language the way you do. It learns statistical patterns of pixels. That single fact explains a lot about both hands and text, and it also explains how the newest models started fixing them.

Models paint patterns, not concepts

A modern image generator is usually a diffusion model. During training it sees billions of images, each paired with a text description. It learns to start from random noise and gradually denoise it into a picture that matches the words. What it is really learning is which pixel arrangements tend to go with which words.

Crucially, there is no symbolic understanding underneath. The model does not hold an idea that a hand is a rigid structure with exactly five digits attached to a palm, or that a word is an ordered string of specific letters. It only knows what these things tend to look like across its training data. When you ask for something, it produces a plausible average of everything it has seen, not a reasoned construction. That works beautifully for textures, lighting and faces. It works badly when the correct answer depends on exact counting or exact spelling.

Why hands are uniquely hard

Hands combine three problems at once. First, they are wildly variable: a hand can be a fist, a wave, a pointing finger, fingers interlaced, or mostly hidden behind an object. The same hand looks like dozens of different shapes, so the model struggles to learn one consistent template.

Second, hands are usually small and not the focus of a photo. As Stability AI has noted, within image datasets hands appear far less clearly than faces, so there are fewer good close-up examples to learn from. Third, the model has no real grasp of the three-dimensional anatomy underneath. It knows how a hand looks, not how it is built, so it has no internal rule stopping it at five fingers. The result was the infamous extra digits. Britannica notes that teeth and ears suffer for the same reasons, being small and highly variable too.

Mitigation came partly from better data. Midjourney shipped an update in March 2023 that prioritized clearer hand images and downweighted obscured ones, and hands improved noticeably, though they were not instantly perfect.

Why readable text was even worse

Text is, if anything, less forgiving than hands. A picture of a hand can be slightly wrong and still read as a hand. A word that is slightly wrong is just misspelled, and your eye catches it immediately. Spelling has no margin for the kind of plausible averaging that diffusion models do so well.

There is also an architectural reason. Many early text-to-image systems pair a language encoder, which turns your prompt into an internal numeric representation, with an image decoder that was trained to make attractive pictures, not to spell. There was no dedicated language decoder converting meaning back into exact letters. So the model treated letters as decorative shapes, painting something that looked like text in the right place without committing to specific glyphs in the right order. Tokenization made it worse, because the model often handles fragments of words rather than clean letter-by-letter spelling. Researchers even found that DALL-E's gibberish was internally consistent: feed the nonsense back in and it sometimes mapped to real concepts.

How newer models got better

Progress on both fronts came mainly from scale, cleaner data, and tighter coupling between language and image generation. Bigger models trained on more varied, better-captioned images saw words and hands in far more contexts, which improved their internal patterns.

Text specifically improved a lot between roughly 2023 and 2025. DALL-E 3 could render short words and phrases that earlier models mangled, and systems known for typography, along with newer general models, now produce legible signage and short captions much of the time. It is honest to say the trend is real but not solved: long passages, unusual fonts, and non-Latin scripts still trip models up, and you should always proofread any text an AI puts into an image.

There is a practical lesson for ecommerce here. For product photos, the most reliable approach is to keep the camera honest and let AI handle the parts it is genuinely good at, like clean backgrounds, consistent white backgrounds and square framing, rather than asking a generator to invent text on packaging or fix the shape of a hand holding your product. Tools like Renderivo lean into that strength: real product, tidy presentation, no hallucinated labels.