I have been thinking about this for ages, but there are two problems:
One is that, like all data, metadata is usually what is missing.
Two, the desire to do it.
Machine learning is at its best when presented with a good data set that is tagged, where relationships and patterns can be deduced, so that meaningful output can be generated.
My feeling is close to Adam’s, in that the information should be apparent from scanning the application itself. Every menu item, every button, every function, needs a structured way to declare itself to the help system. Notwithstanding that writing a manual for your own SW makes for much better code (because you plainly see the absurdities you have made for yourself), this form of declaration could be easier to write and maintain. It might even be made compulsory in the App Store eventually.
I envisage a table of entries declaring properties such as purpose, starting conditions, appearance in the UI, the transformation applied, reversibility, version limitations, etc. ML specialists can better devise ways to design this metadata for the model they use. There are many possibilities. The primary goal, from the user’s POV, is always purpose and meaning.
Awards were won in the past for the quality of Apple’s documentation, but I see no desire to even attempt this level of commitment today. Ex-Apple people have often commented of understaffing of SW development, and I think the corporation no longer institutionally cares about this, seeing the state of their developer documentation.
Yes, superior help and discoverability would be a valuable asset for burnishing the value of the platforms, but does a $4T corporation even register this factor? The idealist visions of the founders are long gone.