For all the recent innovations in imaging, functional brain mapping has no methodology inferentially stronger than the one that launched it, now more than a century ago. And though the look may be sharper, modern lesion-deficit mapping is cut from the same cloth as the old "overlap" techniques, ill-fitted in the mass-univariate style of fMRI and VBM. So enduring a methodology cannot be wrong, it is temptingly argued, but such an appeal is only plausible where errors are inconsistent. Examining the largest published series of images of acute brain injury in the context of stroke, we have recently shown that mass-univariate lesion-deficit mapping is consistently distorted by the architecture of the underlying vascular tree. This distortion is substantial when the lesion-deficit association is implausibly simple, and unquantifiably greater when it approaches the complexity likely to obtain in reality. No historical lesion-deficit map can therefore be trusted, and no future map ought to be traced in the old, mass-univariate way. We must instead construct models that capture the pattern of damage across the brain in its full complexity, inevitably within a high-dimensional multivariate framework that relies on inferential tools drawn from the realm of machine learning and demands datasets on a scale much larger than is the custom in the field.