I wouldn't consider generating descriptions a "solved problem". There has been progress with AI, but part of the issue is that image classifiers are trained to

I wouldn't consider generating descriptions a "solved problem". There has been progress with AI, but part of the issue is that image classifiers are trained to describe the objects in the image, but not necessarily the feeling or meaning of the image. If you have Instagram, scroll through the alt text on the images some time. You'll see descriptions like "May contain: Person standing outside". That may be an accurate description of the picture, but it may be a poor substitute for what a sighted person sees -- someone showing off their new sweater, or posing in front of an interesting wall, or an aesthetic splash of colours. Of course, maybe someone who is fully blind doesn't care too much about a splash of colours, but the point is that describing an image is more than just about identifying the individual objects visible in the image. Because when we see a picture of an adorable puppy, we get more out of that image than just "dog sitting on floor". As of now, human-generated alt text is still the best solution in most cases, even if we are definitely making progress with AI. It's hard enough for humans to describe the aesthetics of an image, let alone a machine.