The light reflected from an object depends not only on the surface properties of this object but also on the illuminant. The same is true for the excitations of the photoreceptors, which serve as the basis for the perceived colour. However, our visual system has the ability to perceive constant surface colours despite changes in illumination. The average chromaticity of the retinal image of a scene depends on the illumination, and thus might be used by the visual system to estimate the illumination and to modulate the correction that subserves colour constancy. But this measure is not sufficient: a reddish scene under white light can produce the same mean stimulation as a neutral scene in red light. Higher order scene statistics-for example, the correlation between redness and luminance within the image-allow these cases to be distinguished. Here we report that the human visual system does exploit such a statistic when estimating the illuminant, and gives it a weight that is statistically appropriate for the natural environment.