Because then everyone dies when the computer fails. Autopilots regularly fail and expect the pilot to take over
I think this depends on your definition of "fail". As far as I know true computer failures where the machine just goes crazy and tries to crash the plane are non-existent. What happens more regularly is the autopilot sees that something weird is happening and chooses to disengage itself - presumably an autopilot program could be written that never disengages and always does the best it can to fly the plane, unless deliberately disengaged.
This is particularly problematic when sensors fail, as they did in AF447, and the computer doesn't know what's going on any more.
No, this is irrelevant. If the planes sensors completely fail then the pilot doesn't know what's going on either, and the plane is probably doomed no matter what. In normal operation these planes are flying in a very small speed corridor between disintegration and stalling. If you don't know how fast your going a stall or overspeed is pretty much inevitable, and if you don't know how high you are even basic visibility problems can cause a crash into the surface. Neither human nor computer can succeed in such a situation.