The core concerns I have with the ARIA technologies:

 * They are relatively complex to author, which means that in general they
   won't be widely used even when they should be, and may even be mostly
   misused when they are used at all (like longdesc), resulting in their
   being mostly useless for their intended purpose.

 * They are not significantly easier to implement or use than new HTML5
   widgets, so it isn't clear why one would implement the ARIA
   technologies instead of just immediately implementing the HTML5 ones.

 * Implementing ARIA roles as a temporary solution to the lack of more
   expressive widgets is expensive, since it doubles the number of
   widget-related features a browser has to support, possibly for decades
   to come, resulting in a much higher cost to implementors (for instance
   in regression testing, bug fixing, documentation, security reviews).

 * They are no more extensible than the HTML5 widgets. To do new widget
   types we'll need new ARIA identifiers and implementation rules, just
   like we would with new HTML widgets.