The increased need for multiple statistical comparisons under conditions of non-independence in bioinformatics applications, such as DNA microarray data analysis, has led to the development of alternatives to the conventional Bonferroni correction for adjusting P-values. The use of the false discovery rate (FDR), in particular, has grown considerably. However, the calculation of the FDR frequently depends on drawing random samples from a population, and inappropriate sampling will result in a bias in the calculated FDR. In this work, we demonstrate a bias due to incorrect random sampling in the widely used GO::Term Finder package. Both T(2) and permutation tests are used to confirm the bias for a test set of data, which leads to an overestimation of the FDR of about 10. A simple fix to the random sampling method is proposed to remove the bias.