Gregory's Server
I myself wouldn't trust any AI with writing alt-texts for the pictures I use in my posts. These are almost exclusively pictures taken in virtual worlds.
First of all, an AI may be halfway capable of recognising virtual objects with exist in the real world just the same, but not perfectly so. As for objects which are exclusive to this kind of virtual worlds or even this particular world, objects which do not exist in real life, an AI won't be able to identify them.
Also, sometimes it matters where a picture was taken. An AI may be able to recognise well-known real-life places, but obscure virtual places, not so much, even less so if there are several such places which look almost or completely identical.
Lastly, even if an AI appeared that is fully capable of correctly describing what can be seen in pictures from virtual worlds, and by that I mean down to even automatically and accurately recognising the creators of objects in case this should matter, even then, the AI would not necessarily be able to know the context and thus decide what is important to mention, what is essential and what is unnecessary. Depending on the context, the very same picture can require different alt-texts, that is, unless people come and demand fully-detailed descriptions of everything that can be seen in a picture, for maybe it could matter to someone.
In case you've got enough patience to read a 6,000+-word rant, I've written a post about alt-text on pictures from within virtual worlds just recently. You can read it here.
First of all, an AI may be halfway capable of recognising virtual objects with exist in the real world just the same, but not perfectly so. As for objects which are exclusive to this kind of virtual worlds or even this particular world, objects which do not exist in real life, an AI won't be able to identify them.
Also, sometimes it matters where a picture was taken. An AI may be able to recognise well-known real-life places, but obscure virtual places, not so much, even less so if there are several such places which look almost or completely identical.
Lastly, even if an AI appeared that is fully capable of correctly describing what can be seen in pictures from virtual worlds, and by that I mean down to even automatically and accurately recognising the creators of objects in case this should matter, even then, the AI would not necessarily be able to know the context and thus decide what is important to mention, what is essential and what is unnecessary. Depending on the context, the very same picture can require different alt-texts, that is, unless people come and demand fully-detailed descriptions of everything that can be seen in a picture, for maybe it could matter to someone.
In case you've got enough patience to read a 6,000+-word rant, I've written a post about alt-text on pictures from within virtual worlds just recently. You can read it here.