There's a steep learning curve full of false summits, for sure, but once you grasp it, there is always an elegant way of solving a problem using it. Sometimes there's even a fast way!
Use of it in the wild, though, leaves a lot to be desired. The false summits lead people to write bad code, the bad code gives it a bad rep, the bad rep leads people to dislike using it, people's dislike leads to then writing bad code... and so on.
XML is not a good form for a transformation language. It's not the quality of the code written in it or anything like that.
Take XSLT and change the syntax, and sure, it might be great. You can't ignore the terrible, terrible choice of making the transformation itself XML. It's completely unreadable and horrific to write.
If most people agree a representation is terrible, it's a bad choice for anything serious, because suddenly only a subset of your team or contributors can work on it effectively.
There is a reason XML isn't used as syntax for programming languages. Anyone claiming that an `if` construct, for example, is well represented by a mess of XML tags has a different brain to me and the vast majority of people I've met.
XSLT was meant to be a general-purpose tool. Having a syntax that the vast majority of people find incredibly hard to work with is clearly a huge problem with the spec.
I mean, sure, it could do the job, but did anyone ever sit down at a file of XSLT and think "yes, this is the best way this information about transforming this XML could be represented"?
JSON is even less suitable for this. It's horrible to read this stuff, horrible to write.