The reasoning is along the lines that there are examples of applications where users do not know how to provide alternate text, such as Flickr, and it has been observed that systems make up for this by generating alternate text from the image's meta data.