Because the filter is used across the entire image, filters learn to detect the features regardless of their position within the image. This turns out to be really useful as it gives us translation invariance, which means we can detect something important regardless of its orientation in the overall image.

Thinking back to MNIST, it's easy to imagine that we might want to detect the loop of a 9, regardless of where it lands in the photo. Thinking ahead, imagine a classifier that classifies pictures as either those of a cat, or a car. It's easy to imagine a set of filters that can detect something as intricate as a car tire. It would be useful to detect that tire regardless of where the car's orientation is in the image, as the presence of something like a tire strongly indicates the image isn't a cat (unless the image is of a cat driving a car).