We propose a method to utilize the ``hard to describe'' but ``easy to verify'' property of unusual events without building explicit models of normal events. One can compare each event with all other events observed to determine how many similar events exist. If an event is normal, there should be many similar events in this large data set. If there are no similar events we consider this unusual: although the event is unknown, it is different from the others. Thus, detecting unusual events in a large data set does not require modeling normal events, but rather the ability to compare two events and measure their similarity.