I don't think that's accurate. All of these models are Markov chains, and Markov chains are short time horizon. They work by knowing what words follow each other. They are not suited to things like subplots and character arcs. What they can do is parrot existing stuff - the product it creates is, by definition, derivative. I just asked it to "tell me a story about a dragon" (my daughter's current favorite). It told me a boring story about a dragon where it clearly knew what dragons were and that they were supposed to slay dragons. I asked it to "tell me a story about a unikitty pegasus" (my daughter's favorite from a couple years ago) and while it had no idea what a unikitty pegasus was, it deduced that it should be full of Care Bears/My Little Pony tropes. When I told it to "tell me about that time an earthworm went to Vegas" it went

I'm sorry, but I don't have any information about an earthworm going to Vegas.

At least Cleverbot gave me "I've never been to Vegas" because Cleverbot is Markoving closer to human conversation.

The task here is "draw a mouse." AI can definitely draw a mouse. "Draw Mickey Mouse" will give you Mickey Mouse. "Create a cartoon character as beloved as Mickey Mouse that the world has never seen before" is a crapshoot that you do not achieve through stochastic processes. We're talking about the difference between "draw a mouse" and "draw Mickey" where "Mickey" doesn't exist yet. Now - will a chatbot accidentally discover "Mickey?"

On the flip side, novels pay so poorly these days that the chatbots might be the only things still writing them.