Alternatively, it’s built into the platform. So when someone uploads an image to Lemmy a local AI model does the description.

Edit: Then it could even be marked as AI generated and people could choose to be exposed to it or not.

Really hard to tell. I mean there are situations in which people think they're doing someone a favour. But they're really not. Upside of doing it individually is: affected people get to pick the model they like best. And they can prompt it however they like. Depends a bit on your expertise on the matter if your pre-generated stuff is on the same level or more a disservice. Upside of pre-generating it once is: maybe a bit less CO2 in the atmosphere and a few less trees killed. But that certainly depends on how many people read those descriptions. If there's just 2 people with screenreaders out there, who don't even click on all the images, you might very well be wasting compute. And have a negative balance on the environment.