Enough people have already commented on what is wrong with Google's AutoLink, most of which I agree 100% with, so I won't delve into that.  Instead, I'll write about how to fix it.

Opt-out solutions (i.e. 'noautolink' meta tag) are useless because they still create chores for content providers where there were none before.  Google's 'rel=nofollow' was a borderline case.  AutoLink is not because the effect of not opting out is greater and more direct.

Short of scrapping AutoLink, a solution that allows content-providers to opt-in is needed.  One such solution to AutoLink only anchor tags with recognizable identifier in the 'class' attribute.  For example, content producers who opt-in can markup following content:

123 ABC St., San Jose, CA
ISBN 0123456789
Harry Potter

like this:

<a class='address'>123 ABC St., San Jose, CA</a>
<a class='isbn'>ISBN 0123456789</a>
<a class='book'>Harry Potter</a>

AutoLink should link the content only IF the class attribute contains one or more keywords (i.e. isbn, ssn, gps, person) it can handle.  There should also be a generic autolink class so content authors can markup without knowing which class to markup a text fragment with.  So above could have been written like this:

<a class='autolink'>123 ABC St., San Jose, CA</a>
<a class='autolink'>ISBN 0123456789</a>
<a class='autolink'>Harry Potter</a>

If the anchor tag is not appropriate, then I think 'span' tag can be used with minimal visual side-effects.