Background: The aims were to evaluate the accuracy of self-report of past-month cannabis use in a representative sample of French military staff members and to evaluate the scale of the prevarication bias. Method: Data from three cross-sectional surveys conducted between 2005 and 2008 (n = 3493) were used. The characteristics of self-report (sensitivity, specificity, positive predictive value and negative predictive value) were computed using tetrahydrocannabinol detection in urine as the reference. Results: The prevalence for past-month cannabis use was 16.1% and for positive testing was 13.4%. The discriminant power of self-report was good, with an area under the receiver operating characteristics curve 0.90. Specificity (94.5%) and negative predictive values (97.8%) were good, but sensitivity (85.7%) and positive predictive values (70.4%) were lower. The lowest sensitivity values were observed in the higher categories of personnel and in the Navy, which could reflect some prevarication in these sub-populations who might believe they were more exposed to sanctions if detected. Conclusions: Despite certain limitations of urine analysis as a reference, because of its poor detection of occasional users, our study is in favour of good accuracy of self-reported data on cannabis use, even among the military. However, our results, derived from a population study, do not enable any assumptions on the validity of self-reported data collected during individual testing procedures for the purpose of improving occupational safety.