State of the art natural language generation using deep generative models can produce text that sounds somewhat coherent and has decent grammar, but it lacks the clarity, flow, and purpose that a human writer can produce. I don't believe this points to an inherent flaw in the AI as these systems are very simple. The GPT-2 produced by OpenAI was trained using a language model that predicts the next word. That's it. It was large in the number of parameters, but simple in its function. I can imagine much more complex approaches using today's technology, but they still require years of engineering to train and perfect.
So, I believe that even without further breakthroughs, much better language generation systems can be built. But these systems will still lack the creativity of a human and require several years of well-funded development. Truly human-level writing will not happen in the next 20 years. It's too difficult to predict beyond that horizon, but our minds appear to be physical systems that will eventually be replicated.