Monte Carlo (MC) based simulations of photon transport in living tissues have become the "gold standard" technique in biomedical optics. Three-dimensional (3D) voxel-based images are the natural way to represent human (and animal) tissues. It is generally believed that the combination of 3D images and MC based algorithms allows one to produce the most realistic models of photon propagation. In the present work, it is shown that this approach may lead to large errors in the MC data due to the "roughness" of the geometrical boundaries generated by the presence of the voxels. In particular, the computed intensity of the light detected on the tissue surface of a simple cubic tissue phantom may display errors from -80% to 120%. It is also shown that these errors depend in a complex manner on optical and geometrical parameters such as the interoptode distance, scattering coefficient, refractive index, etc. and on the degree of voxelization ("roughness") of the boundaries. It is concluded that if one wants to perform reliable 3D Monte Carlo simulations on complex geometries, such as human brain, skin or trabecular bone, it is necessary to introduce boundary meshing techniques or other equivalent procedures in the MC code to eliminate the deleterious effect of voxelization.