It is often the case that negative voxels do not replicate very well -- that is, if you have multiple imaging runs with the same paradigm in the same subject, the positive voxels will replicate pretty well, but each run will give mostly non-overlapping sets of negative voxels. If you have multiple runs, you may want to check this.
Highly repeatable negative results can be interpreted as Carl suggested. They clearly mean something. One hypothesis that I've heard put forward is that negative activations can be caused by internal speech/though that is interrupted by task performance/attention. IANANS, YMMV.
bob cox