I'm not sure if I actually accept the premise, but I'll stab at it anyways.

There are some parts of our brains that seem to handle certain specific visual cognitive tasks. Apparently recognizing faces is one?

Letters and hands I think occupy a similar niche in that we are very unforgiving of incorrectness. I expect hands because of the importance of communication (similarity to the whites of our eyes), and letters because we're so trained to read that it's very hard to ignore or forgive errors there.

It might not be that AI is actually especially bad at letters and hands... it might be that our brains are hard wired to be exceptionally critical of letters and hands. AI famously lights scenes impossibly, but most people don't notice.

As for dreams, I think if you think about hands and letters both encoding salient information even "at idle"... generating what that information ought to be in the context of a dream might be challenging.

So, hands and letters just might be cursed with having very human-centric meanings for correctness that we are highly attuned to, so generating these things bears an exceptionally high burden. Brain avoids it in dreams and ai struggles too.