When the distribution of input data changes between training and real-world use, causing models to fail.