My biggest fear is that accidents of this type will result in the engineers running back to their ML to add another case to the training data. We'll end up pla

My biggest fear is that accidents of this type will result in the engineers running back to their ML to add another case to the training data. We'll end up playing a game of whack-a-mole (literally?) as new special cases come up. Will training the system to recognize a pedestrian pushing a bicycle enable it to recognize someone riding a mountain bike? Consider why the auto-braking system wasn't enabled. It's because the system can't identify which elements in the environment are harmless, causing too many false positive braking events. The opposite problem is no easier to solve.