Large structure in tomographic images of the core-mantle boundary (CMB) found by fitting low-degree spherical harmonic models to International Seismological Centre (ISC) travel time data by linear least squares may be artifacts of the fitting procedure, rather than necessary features of Earth. Two numerical experiments show that gaps in the data coverage and the use of low-degree models tend to yield spurious topography. We fit a large set of summary PKP and PcP rays for 23 years of ISC arrival times (for events relocated with the IASP91 model) to obtain least squares CMB ''reference models'' of maximum degree 4, 8, and 12. Least squares CMB models derived from subsets of the data with new artificial gaps comparable in size to the gaps in the original data set can have significantly different structure than the reference models, both globally and within the artificial gaps. In spite of these differences, the correlation of models with different sized gaps is high: correlation is not a good measure of model similarity nor of reliability of the fitted models in this problem. We generated synthetic data from the degree 12 reference model rotated relative to the source-receiver frame. Least squares fits of low-degree models to the synthetic data show significant features not present in the reference model, in approximately the same place relative to the gaps in the original observations. The spurious features of the models with new gaps and with rotated synthetic data are qualitatively indistinguishable from large features in the reference models. ¿ American Geophysical Union 1993