You need to understand the initial use case for XML. It was invented for document-oriented markup languages like HTML, MathML, Docbook etc. You can definitely write XHTML by hand, and a JSON-based syntax for the same kind of documents (which mixed content and so on) would be a lot harder to read and write.

My understanding is that XML was derived from document-oriented SGML, to beat SGML into a form that would work well with XSL and XPath.

But I'd like to point out that the way SGML-derived markup distinguishes attributes and child nodes is entirely arbitrary. You could as easily make attributes child nodes - it's all in how you interpret what's written. Likewise, you can "convert" SGML to JSON (or YAML or S-expr or whatever) very easily, bearing in mind that attributes and child nodes sit in the same space with each other - a well-formed XHTML document, for example, can be re-expressed in JSON without ambiguity, since tags have a well-specified, unambiguous list of allowed children and attributes - just give text nodes the name "text" and you're golden.