A psychoacoustic experiment was conducted to evaluate the audibility of spectral detail of head-related transfer functions (HRTFs) at high frequency. The magnitude of individual HRTFs above 5 kHz was smoothed by a moving frequency window with different bandwidths and subsequently used to create virtual sound images in headphone reproduction. Results with six subjects demonstrated that the ipsilateral and contralateral HRTFs in both the horizontal and lateral planes could be smoothed only with a bandwidth of 2.0 ERB (equivalent rectangular bandwidth) and 3.5 ERB, respectively, without introducing audible artifacts. Both magnitudes of the binaural HRTFs above 5 kHz could also be smoothed with a bandwidth of 2.0 ERB for the ipsilateral ear and 3.5 ERB for the contralateral ear, respectively; whereas in the median plane, both magnitudes of the binaural HRTFs could be smoothed with a bandwidth of 2.0 ERB without introducing audible artifacts. Therefore the spectral detail of HRTFs at high frequency is inaudible. A binaural loudness model was further used to analyze the experimental results. The audibility of the smoothing artifacts of HRTFs was indicated to depend on the difference between the (monaural and binaural) loudness spectra caused by the original and smoothed HRTFs. Consequently, the result was both spatial direction and individual-dependent. In practical signal processing of the virtual auditory display, it is unnecessary to use complex signal models to represent the spectral detail in HRTF magnitude at high frequency.