I have doubts on Davids explanation too.
Imagine a "perfect" ADC without DNL and noise and apply a linear voltage ramp to the input. Then the quanization noise should manifest with a sawtooth-waveform. Amplitude would be 50% of quantization intervall, fundamental frequency of the sawtooth would be determined by slewrate of inputsignal divided by the quantisation intervall. The spectrum of quantizaiton noise would be the line spectrum of a sawtooth waveform.
Adding DNL to this ADC should "modulate" or "jitter" the fundametal frequency. Also in real world situations the quantization noise can be "coloured", though the exact noise spectrum will indeed be very hard to predict.