AI upscaling takes a small, low-resolution image and produces a larger, sharper version, but the sharp detail in the output was not recovered from the original—it was invented by a neural network that learned what sharp detail usually looks like in similar images. Before AI, upscaling used mathematical interpolation: nearest neighbor copies each pixel to fill the larger grid, so a 100×100 image scaled to 400×400 repeats each pixel in a 4×4 block, producing blocky, pixelated results useful for pixel art but terrible for photographs; bicubic interpolation blends neighboring pixels using a weighted average, producing smoother but blurry results in which fine detail becomes a muddy smear. Both methods cannot add information that is not there: a 200×200 photo contains roughly 40,000 pixels of information, while scaling it to 800×800 requires 640,000 pixels, and the missing 600,000 pixels are filled with mathematical averages of their neighbors, producing blur. AI upscaling instead predicts missing high-resolution detail based on patterns learned from millions of training images, which are systematically degraded by shrinking, adding noise, applying blur, and compressing with JPEG artifacts in random combinations, then reconstructed and graded against the original high-resolution version.