In a genetic selection designed to isolate Escherichia coli mutations that increase expression of the IS 10 transposase gene (tnp), we unexpectedly obtained viable mutants defective in translation initiation factor 3 (IF3). Several lines of evidence led us to conclude that transposase expression, per se, was not increased. Rather, these mutations appear to increase expression of the tnp'-'lacZ gene fusions used in this screen, by increasing translation initiation at downstream, atypical initiation codons. To test this hypothesis we undertook a systematic analysis of start codon requirements and measured the effects of IF3 mutations on initiation from various start codons. Beginning with an efficient translation initiation site, we varied the AUG start codon to all possible codons that differed from AUG by one nucleotide. These potential start codons fall into distinct classes with regard to translation efficiency in vivo: Class I codons (AUG, GUG, and UUG) support efficient translation; Class IIA codons (CUG, AUU, AUC, AUA, and ACG) support translation at levels only 1-3% that of AUG; and Class IIB codons (AGG and AAG) permit levels of translation too low for reliable quantification, importantly, the IF3 mutations had no effect on translation from Class I codons, but they increased translation from Class II codons 3-5-fold, and this same effect was seen in other gene contexts. Therefore, IF3 is generally able to discriminate between efficient and inefficient codons in vivo, consistent with earlier in vitro observations. We discuss these observations as they relate to IF3 autoregulation and the mechanism of IF3 function.