As several others have pointed out, realism of these models will continue to improve, and will soon be economically useful for producing beautiful or functional artifacts - however prompt adherence (getting what you want or intend) of the models is growing much more slowly.
However I think we have a long ways to go before we'll see a decent "AI Film" that tells a compelling story - and this has nothing to do with some sort of naturalistic fallacy that appeals to some innate nature of humans!
It comes down to the dataset and the limits of human creators in their ability to communicate their process. Image-Text and Video-Text pairs are mostly labeled by semi-skilled humans who describe what they see in detail. They are, for the most part, very good at capturing the obvious salient features of an image or a video. "reflections of the neon lights glisten in the sidewalk". However, what you see in a movie scene is the sum total of dozens if not hundreds of influences, large and subtle. Choices made by the actors, camera operators, lighting designers, sound designers, costuming, makeup, editors, etc... Most people are not trained to recognize these choices at all, or might not even be aware that there are choices to make. We (simply) see "Joaqin Phoenix is making awkward small-talk in the elevator with other office workers".
So much of what we experience processes on subconscious and emotional and purely sensory levels, we don't elevate those lower-level qualia to our higher-brain's awareness and label them with vocabulary without intentional training (such as tasting wine, coffee, beer, etc - developing a palate is an act of sensory-vocabulary alignment).
However, despite not raising these things to our intentional awareness, it has an influence on us -- often the desired impact of the person who made that choice in the first place. The overall effect of all of these intentional choices makes things 'feel right'.
There's no fundamental reason AI can't produce an output that has the same effect as those choices, however finding each little choice is like a needle in a haystack. Accurate labeling of the training data tells the AI where to look -- but the people labeling the data are probably not well-versed in all of the little intentional choices that can be made when creating a piece of video-media.
Beyond the issue of the labeling folks being trained in the art itself, there's the problem too of the artists themselves not being able to fully articulate their (numerous, little, snowflake-into-avalanche) choices - or simply not articulating it even if they could. Ask Jackson Pollock about paint viscosity and you'll learn a great deal, but ask about abstract painting composition and there's this ineffable gap that language seems ill-suited to cross. The painter paints what they feel, and they hope that feeling is conveyed to the viewer - but you'd be hard pressed to recreate "Autumn Rhythm (Number 30)" if you had to transmit the information via language and hope they interpreted it correctly. Art is simultaneously vague and specific!
So, to sum up the problem of conveying your intent to the model:
- The training data labels capture obvious or salient features, but not choices only visible to the trained eye
- The material itself is created by human artists who might not even be able to explain all of their choices in words
- You the prompter might not have the vocabulary that captures succinctly and specifically the intended effect
- The end result will necessarily be not quite what you imagined in your mind's eye as a result of all of this missing information
You can still get good results if you tell it to copy something, because the label "Tarantino" captures a lot of detail, even all the little things you and the training data would never have labeled in words. But it won't be yours and - until we have an army of trained artists providing precise descriptions for training data in their area of expertise, and you know how to speak those artists' language - it can't be yours.