Label errors that depend on the features of individual examples, making them harder to detect than random noise.