In this paper we investigate the impact of inconsistency in manual annotations when they are used to train automatic models for 3D facial landmark localization. We start by showing that it is possible to objectively measure the consistency of annotations in a database, provided that it contains replicates (i.e. repeated scans from the same person). Applying such measure to the widely used FRGC database we find that manual annotations currently available are suboptimal and can strongly impair the accuracy of automatic models learnt therefrom. To address this issue, we present a simple algorithm to automatically correct a set of annotations and show that it can help to significantly improve the accuracy of the models in terms of landmark localization errors. This improvement is observed even when errors are measured with respect to the original (not corrected) annotations. However, we also show that if errors are computed against an alternative set of manual annotations with higher consistency, the accuracy of the models constructed using the corrections from the presented algorithm tends to converge to the one achieved by building the models on the alternative,more consistent set.