>I don't know why anyone would assume that, given how poorly it works.

You're tough. Every type of computer generated visuals have some cases that won't work well, given that there are so many semantic choices: anti-aliasing, dithering, textures, forms, compete for the information of each pixel. Each pixel can encode any subset of these features. We rely on the human brain to parse it and decode it meaningfully, and even then, it doesn't always work. If you've tried yourself at pixel art, you'll understand quickly (you can also look at the research on upscaling algorithms). So the result is very far from an actual product, but it's really impressive.