“Generate an image of a man holding a red helium balloon in his left hand”
This was the prompt I asked OpenAI’s Image Generator DALL-E 2, 2 years ago. The result? A perfect shiny red helium balloon on a string with a man in a blue shirt holding it with two… things.
On a closer inspection I could make up that it, in fact, tried to make two hands, but it had failed miserably. The fingers were merging into eachother and his hands were also connected while holding the string of the balloon. Poor guy.
When I tried this again, the same thing happened. A beautiful image of a man holding a red shiny balloon but it again had failed at generating a realistic image of the hands. So why is it that these AI-Image Generators oftentimes struggle so much with generating hands and fingers?
Back in 2022 when these models were released, people were amazed by the images that these generators were spitting out. The images were so advanced that in August of 2022, the Image Generator Model from the company Midjourney won an art contest at a state fair with one of its images.
But users quickly started to notice a recurring bug. Everytime when a prompt included people, the AI tools couldn’t draw hands. Hands with 7 fingers, hands that appeared to be floating, unattached to the human body or, in my case, hands that were fused together at the wrists. But why?
The simple answer? Its hard to draw hands! Just like beginning (human) artists, AI struggles with hands because a hand is a very complex part of our body. It has multiple elements of varying shapes and sizes. In addition to that, its structure is incredibly intricate. Hands are built from parts that fit together with perfect precision. Fingers, palms, joints, and tendons all connect in fixed patterns. To draw them well, you must study how these parts move and align.
The case with AI models is that they learn by finding patterns in data. They do not actually understand structure as humans do. A human artists learns through observation and reasoning, the models learn through repetition in data. In addition to that, to reallisticly capture how the hand deforms during various hand movements, algorithms need to understand how our joints function and the range of motion each of them capture. Which is increadibly hard and tedious to train.
However, as time passes there are definitly improvements being made. Images with 6 fingers or hands fused into each other are less common now. That is because new models use larger datasets with millions of hand examples. They learn cleaner shapes, smoother joints, and more natural poses. The improvement isn’t magic. It comes from better training methods and higher quality references. Still, the model doesn’t understand anatomy. It predicts what looks right, not what is right. You can see progress, but the small mistakes remind you that pattern recognition is not comprehension.
And about that image I mentioned earlier with the prompt of a man holding a red balloon, here it is:
Sources:
Avenga. (2025, 18 juni). Why Generative AI Models Fail at Creating Human Hands – Avenga. https://www.avenga.com/magazine/generative-ai-models-fail-at-creating-human-hands/#:~:text=So%2C%20Why%20Do%20Image%20Generators,may%20take%20it%20for%20granted.
Matthias, & Meg. (2023, 25 augustus). Why does AI art screw up hands and fingers? | Explanation, Tools, & Facts. Encyclopedia Britannica. https://www.britannica.com/topic/Why-does-AI-art-screw-up-hands-and-fingers-2230501
Great post! I had a similar experience a few years ago when I tried generating images of athletes. If I remember correctly, it was with the Canva AI image generator. Everything looked great (the lighting, motion, and even small details). Except for the hands, which looked like something from a sci-fi movie – 3 fingers or hands blended together. I tried to use prompts to fix it, but at that time, I was not really successful.
What you mentioned makes me think that these tools mirror human learning in an interesting way: they can imitate expertise but not truly understand it. The progress since 2022 has been incredible, but as you said, they still predict what looks right, not what is right.
Do you think AI will ever move beyond imitation and actually learn structure the way humans do?
Great example to explain this complex problem. The part about pattern recognition and comparing that to comprehension is very interesting. I wonder if the new multimodal models that integrate 3D and motion data might actually be able to overcome this limitation. This example was two years ago, but rate at which AI is has progressed since then already feels like a completely new generation of tools. If AI could learn from actual movement rather than just images, maybe it could start understanding anatomy rather than guessing it. Do you think that’s realistic or still too far away? Have you tried using newer models and seeing whether the outcome is more realistic or better?
I really enjoyed your post! The way you described the “two things” that were supposed to be hands made me laugh because I had almost the same experience. I recently used the Remini AI app to generate a professional LinkedIn photo – it actually looked quite realistic (my friends couldn’t tell it wasn’t taken by a photographer), but the hands were a total disaster. In most versions, the fingers were fused or oddly shaped, so I just cropped them out. What made it even more frustrating was that Remini charges a subscription fee, so I was essentially paying for pictures I couldn’t fully use.
I also agree with the point raised in the earlier comment, these tools can imitate expertise but not really understand it. Your line about AI “predicting what looks right, not what is right” sums that up so well. Maybe integrating 3D motion data could help future models grasp structure more accurately. Do you think we’ll ever reach that point?
Interesting post! I never tried it myself, but it is interesting to see that generating realistic hands is still a major challenge for AI. It really emphasises that while generative AI can be incredibly useful, we need to approach it with caution. It is easy to assume that AI outputs are perfect, but small mistakes like these highlight the limits of these models. Maintaining a critical stance is essential, especially if it is being used for professional or creative purposes.
I also find it interesting how this mirrors human learning. AI can imitate patterns but does not truly understand the underlying structure yet. Do you think this limitation affects how we should use AI in creative projects?
I really liked your post. The way you explained why AI struggles with drawing hands made it easy to understand. It’s funny how things that seem so ordinary to us can show the limits of something that seems so advanced. I also agree with your point about improvement. Even though the results look better now, it still feels like AI is copying rather than really understanding what it creates. I wonder if AI will ever move from recognizing patterns to actually understanding form and structure.