Low degree misses the curve; high degree chases each sample's noise.
Bias² , variance
A rigid model misses the curve the same way on every sample; a flexible one chases each sample’s noise in its own direction. Thin lines fit six noisy samples of the dashed curve, the thick line is their average. Raise the degree and watch the average close in on the curve while the fits fan out and test error turns back up; the samples are illustrative.