As an artist and designer (with admittedly limited AI experience), where I feel AI to be lacking is in its poverty of support for formal descriptors. Content descriptors such as 'dog wearing a hat' are a mostly solved problem. Support for simple formal descriptors such as basic color terms and background/foreground are ok, but things like 'global contrast' (as opposed to foreground background contrast), 'negative shape', 'overlap', 'saturation contrast' etc etc... all these leave the AI models I have played with scratching their heads.

I like how Veo supports camera moves, though I wonder if it clearly recognizes the difference between 'in-camera motion' and 'camera motion' and also things like 'global motion' (e.g. the motion of rain, snow etc).

Obligatory link to Every Frame a Painting, where he talks about motion in Kurosawa: https://www.youtube.com/watch?v=doaQC-S8de8

The abiding issue is that artists (animators, filmmakers etc) have not done an effective job at formalising these attributes or even naming them consistently. Every Frame a Painting does a good job but even he has a tendency to hand wave these attributes.