Three-dimensional reconstruction from electron microscopic (EM) images of isolated macromolecular complexes is being employed by many laboratories. This approach is extremely powerful and continues to improve in resolution. In the absence of stereochemical constraints that can be used to assess the quality of a reconstruction, as exist in X-ray crystallography, several other measures have typically been used. A very useful assessment of quality can be made in the comparison between the projections of the three-dimensional reconstruction and averages generated from classes of images. The main quantitative measure has been that of resolution statistics, typically based upon Fourier shell correlations. We show, using only simulated noise for images, that impressive resolution statistics are generated that can even extend the apparent resolution of the starting model. When truly independent reconstructions are generated starting from different initial models, however, such artefacts are not possible. We also show, using real images of DnaB rings, that in the presence of polymorphism artefactual reconstructions can be generated whose projections match class averages. These averages, however, are themselves artefactual as they involve heterogeneous images. The issues presented here need to be considered when single-particle EM reconstructions are evaluated. (C) 2003 Elsevier Inc. All rights reserved.