Why does AI fail to generate simple ASCII images ?

I saw a post earlier about MineBench. I was impressed to see that the latest models can produce such realistic outputs. Their ability to understand the prompt and make spatial modifications were impressive. But when I asked the models to generate simple ascii images, they failed spectacularly. Prompt: Draw simple ascii image of a person touching his eyes. gemma-4-31b-it O / /|/ / \ (looks like someone hung themselves to me) grok-4.1-thinking (=⌵=) ( x x ) ( ─ ) |||| |||| / \ (=⌵=) ( x