I'm concerned about the opposite possibility. All the effort will go into tweaking models and adjusting their training data until they generate acceptable output, and documentation for human consumption will be an afterthought.
> documentation for human consumption will be an afterthought.
The documentation just changed a layer upwards in the abstraction ladder. Instead of docs for the code that runs on the machines, we'll write docs for the programs we've built that can generate the code that runs on the machines.
