If the internal guidance fails, it's likely not because of a bug in the software, but failure of sensors. If you lose sensors, you lose the information about where you're going anyway and a human operator does you no good.
It all depends which sensors you lose, and what type of failure they experience. If a sensor is installed in such a way that its readings get sign flipped, a human is likely to be able to recognize and correct for the error, whereas an automatic guidance system is likely to fail catastrophically. At the very least, a human is more likely to be able to keep the correct end pointed towards space based on visual reference from cameras alone.
More than that, while we all fly rockets in Orbiter, I think piloting a real rocket would be much more difficult.
Sure. But given the choice between a human stretched to the limits of his abilities but aware of what he's doing and a guidance system that is well within its limits but hopelessly confused by bad data, I'll take the human.
You also have to consider that the handoff from on-board guidance to remote guidance would take at least 1 second for the operator to really start controlling the rocket. And in that 1 second, a lot can go wrong.
In situations where the guidance software is receiving incorrect data from sensors, it is likely to be making control inputs that exacerbate the situation (if a sensor input is sign flipped, control inputs that depend on that sensor are likely to be sign flipped as well). In such situations, during the interval between the end of internal guidance and the beginning of remote guidance, while the operator is not yet doing anything to correct the situation, the on-board guidance is at least no longer acting to make it worse.