In T326041#8599931, @cscott writes,

I think this option (removing the 'alt' attribute) is good and preferable to the alternative (removing just the file extension), as my understanding is that modern screen readers can actually do a limited form of machine vision processing on images without alt tags. It's not as good as human-generated alt info, of course, but they can describe the general image ("yellow rectangle on green background") and even do text-to-speech on any embedded text. I assume (feedback welcome!) that is more helpful than just repeating the filename, and that the screen reader isn't going to do any of that magic if it "thinks" there's a valid alt attribute already present. So in the interest of "honest output" I think suppressing the alt attribute when we don't actually have any human-generated content to put there is the right thing to do.

@Tkinkhorst Thanks for the patch, and sorry for the slow review, but we've gone the route of no longer falling back to the filename for the alt attribute.