Cut a circle of circumference D at N random points, then drop one observation. The arc it lands in tends to be longer than D/N -- the expected length is 2D/(N+1). The act of observing biases the measurement.