The next stage of evolution of the web is not ever fancier graphics and formatting, aimed at eyeballs, but more structured information, aimed at brains.
Making the TeX source available by default makes it substantially easier for automatic tools to arise which can adaptively search and interpret the xxx archive for human scientists. Sure, they aren't here yet, but most mathematicians used TeX long before they used the Internet. Consider how the widespread adoption of TeX has enabled the Physics and Mathematics communities to take advantage of the Internet for scholarly communication much more quickly than our colleagues in other disciplines where text is prepared to less portable standards.
A dumb, but dogged, machine can do amazing things with structured information. I sat in a lab several years ago with a workstation that contained the full text of a moderate sized encyclopedia and a ``scatter gather'' program which iteratively presented me with a list of 10 keywords that it had extracted from the text to automatically group the material into subject areas. I would choose three, and then it would narrow its search. In four iterations, and I had gone from the entire encyclopedia to a small subset of articles arranged according to something very much like the AMS classification scheme for Geometry! Another thing a machine could easily do with TeX source is detect plagiarism, if that really is a legitimate concern. Most of the widely publicized accounts about the rate of scholarly plagiarism are, in fact, based on textual analysis carried out by machines.
At the moment, the TeX source is the richest, most structured form in which the information in our papers exists [outside the minds of the authors and readers]. By making anything less available, we are muffling our scientific expression.
8/17/98