This assumes that rides outnumber drivers so that drivers can be picky. In practice tho, it should be the other way around: drivers should be choosing between a paid ride to "bad" destination, and no ride for some time while they're waiting. You can trivially teach the algorith to give better rides to less choosy drivers first.

Second, if you still have amateur drivers on your platform who do ad-hoc carpooling, chances that they're looking for an one-way drive to the same "bad" destination because they have further non-cab business there. You can teach the algorithm to offer a selection of rides to such drivers for their first ride in a few hours. Or even let them to specify where they want to go. Maybe it's already implemented by someone?

And yes - Uber price is always unpredictable for a rider, so why not just factor in destination desirability into it?