Sure, you can add information until you arrive at a point where a string attribute or text content will be interpreted as a certain data type, but XML itself doesn't have it.
> JSON also has similar higher-level abstraction layers with more metadata.
JSON has all the basic data types built right in, there's no need for more metadata to do simple things. There's a reasonable mapping to basic data types and structures for almost any language.
> The problem is that nobody can agree on which one to use, or even whether to use one at all, and most code that's deserializing JSON in the wild is not going to be able to distinguish metadata from data.
...which is generally fine because of the aforementioned mapping. Your JSON library doesn't have to (and shouldn't) do any magic.
XML (or rather, XDM, which is the appropriate level of abstraction to talk about this) has all those things:
https://www.w3.org/TR/xmlschema-2/#built-in-primitive-dataty...
Nor does it doesn't require an out-of-band schema - you can slap xsi:type on any element. And you can do that without breaking the data model, because namespaces keep data and metadata unambiguously separate, and code can easily deal with the former while being completely oblivious to the latter, unless it needs it.
JSON also has similar higher-level abstraction layers with more metadata. The problem is that nobody can agree on which one to use, or even whether to use one at all, and most code that's deserializing JSON in the wild is not going to be able to distinguish metadata from data.