Because soon you will just ask your computer to build your code after giving it a few parameters. Then you will proof read and sanity check the code and spend more time optimizing and working on the really hard problems.
In Automator, pre-made blocks of code that do different things (filesystem manipulation, data transformations, application control, machine control) can be assembled into a workflow, and configured with options. It is possible that something like that could be given a speech-based UI (rather than keyboard and mouse) and the result is that pre-existing code is assembled and configured. That's a lot more realistic.
Now, machine learning to find bugs or adjust variables... yeah. I can see that.
On the contrary, I think machine learning might be more immediately useful for things like low-level optimization.
For example if I remove a global variable from a source file and another branch adds another reference to it, if it does not occur on the same line git will happily merge the file with no conflict.