Because with real world data like in production in tech there are so many factors to account for. Brittle methods are more susceptible to unexpected changes in the data or unexpected ways in which complex assumptions abut the data fail.