Until there exists a Neural Net proofen to be able to validate any contract specified for a NN to uphold.
After that its unemployment, like the rest of humanity.
Even so, if statements tend to be written by humans, compartmentalised into functions/modules/methods/etc. and plugged together to form whole systems. Microsoft software (presumably) isn't one huge `if/elif/elif/...` in a single `while` loop, even though that would suffice in theory.
Likewise, I imagine Google's software isn't one giant formula, conditioned on inputs and sampled for outputs. Although many of their user-facing applications do contain a black box of this form (e.g. for translation, text-to-speech, etc.), they're only used in very narrow, well-defined tasks, and are surrounded by a huge amount of "normal" code.
But at the end of the day you need something human (or equivalent) to validate it, and as any programmer knows it's a lot easier to understand and validate something you've written than something someone (or something) else has written.
Seeding a NN would be a simple listing of invariants: Headers are at the top of the screen, footers at the bottom, there exists a name field in the page and the datastore; the name field on the page corresponds to the name field in the datastore. Once those invariants are defined, the intermediate product is released to the customers and they directly drive further development of the NN by simply saying "this is not the expected behavior" in some easy drop-down.
I think the biggest hurdle will be that a human's assumptions are going to be more palatable than a NN's assumptions. But once a number of those assumptions are codified (or a statistical preference is determined), this weakness may be effectively mitigated.
This "crowd-sourced" definition of correctness has already been proven - in chatbots. People have been able to coerce chat bots into providing their desired output without touching a line of code.
[1] http://blog.foretellix.com/2017/07/06/where-machine-learning...