We examine the inductive biases of conditional diffusion and devise a rigorous metric for when diffusion models do not behave like denoisers.