I don’t think we’ll see Neural Nets creating entire television shows any time soon, in addition to other classes of drawings/images, for two reasons.
1) During the act of creation, human beings are able to constantly evaluate what looks good according to their own tastes, thus acting in a both backwards and forwards search through the content space. A regular neural network which is orders of magnitude less complex cannot accomplish this or logically reason about it.
2) Filling in certain elements requires higher abstract reasoning, which is more complex and also requires a lot more power. (e.g., logically deducing facts about things which are not immediately visible in the scene)
What I think we’ll see instead are patches, filters, and tools. Not “make me a South Park episode”, but “make me a texture of gravel”. An existing one sharpens up pictures of animu girlz, and likely a NN could be trained to, for example, increase the resolution on old 90′s anime.
https://en.wikipedia.org/wiki/Generative_adversarial_networks
(I do agree that low-level feature generation will be used first)
Oh, it may well get there eventually, but with the slowdown on Moore’s Law, it depends on one’s evaluation of just how powerful and just how inefficient human brains are - and if they’re doing stuff with microtubules and non-trivial quantum stuff, that suggests something more rather than less powerful than previous estimates.
It’s likely that, aside from the requirements for replication (itself non-trivial), the human brain’s computational density isn’t that horrifically poorly tuned, probably reflecting tradeoffs in energy consumption, heat dissipation, latency, durability, and so on.
Currently, I’m forecasting specialized modules on dedicated hardware, but I don’t think we’re going to hit the levels of computational cost/density that Kurzweil and the other Singulatarians predicted, as the cost in engineer manhours per chip is going up, last I checked.
(I think I’ve already mentioned some implications for ethics.)
It does cause the information risks you’re worried about long before being able to create an entire South Park episode, however.

