Human vision has the remarkable property that, over a wide range, changes in the wavelength composition of the source light illuminating a scene result in very little change in the colour of any of the objects. This colour constancy can be explained by the retinex theory, which predicts the colour of a point on any object from a computed relationship between the radiation from that point and the radiation from all the other points in the field of view (Fig. 1). Thus the computations for colour perception occur across large distances in the visual field. It has not been clear, however, whether these long-range interactions take place in the retina or the cortex. Reports that long-range colour interactions can be reproduced binocularly when one band of wavelengths enters one eye and a different band enters the other might seem to establish the cortex as the site of the computation. Many observers, however, see very unsatisfactory colour or no colour at all in this binocular situation, suggesting that the cortex may not be the only site at which the computation is carried out, or even the most important site. We have now tested the role of the cortex in a human subject in whom the nerve fibres connecting cortical areas subserving two separate parts of the visual field had been severed, and find that the cortex is necessary for long-range colour computations.