A simple example is ‘\sum’ which provides no way to capture the expression being summed over - because that’s not necessary for typesetting. That’s not the case in, say, MathML.
Writing MathML is no fun though because mathematical formulae are visually ambiguous and we rely on the context to know how to read them, e.g. does ‘f(x - 1)’ mean function f called with argument x - 1, or does it mean variable f multiplied by x - 1?
Say I have Equation \ref{eq}. Why can't I just say "plot \ref{eq} for x from -6 to 11" and get my graph?
And yes, I know about pgfplots, PSTricks, TikZ etc. But in all those cases, I need to define the same equation twice, in different syntax to boot. It's kind of unsatisfying.
Pretty much for the same reason you cannot press a word and get a pop-up dictionary definition in a paper book.
An often cited example: what is f(x+y) ? Is it function f with x+y as its argument, or constant f multiplied by (x+y) ? TeX gives you no clue.
Or what is this i in your equation? Is it an index variable, or a square root from minus one?
You as a human figure this out by looking at the context and using domain knowledge. So does a "TeX to HTML/MathML converter". It is ultimately built on heuristics, and cannot be otherwise.
That's why I said basically "for the same reason a paper page is not interactive". It was designed this way!
The goal of TeX was to generate beautiful printed page. The need for semantic structure was not anticipated. To do semantics you need a "semantic version of MathML", or a language used by Wolfram's product, etc.