A skier might conclude that had he bent his knees on a certain slope he would not have fallen. The skier can learn from this. It is important, if we are to use counterfactuals for learning that we can recognize that they are sometimes false. In our first example we can imagine the response that ``If there were another car it would have been visible in time for me to avoid it''. This new counterfactual can also be true or false.
The performance of learning algorithms improves when they have more examples. Counterfactuals are one way to collect more examples than can be found by direct experience. Often it is better to imagine a data point than to experience it.