We report here the use of the mutual information theory for the certification of annotated rice coding sequences of both GenBank and TIGR databases. Considering coding sequences larger than 600 bp, we successfully screened out genes with aberrant compositional features. We found that they represent about 10% of both datasets after cleaning for gene redundancy. Most of the rejected accessions showed a different trend in GC3% vs GC2% plot compared to the set of accessions that have been published in international journals. This suggests the existence of a bias in the pattern recognition algorithms used by gene prediction programs.