In 1995 the International Agency for Research on Cancer (IARC) completed a study that involved nuclear workers from facilities in the USA, UK and Canada. The only significant, though weak, dose-related associations found were for leukaemia and multiple myeloma. The results for the Canadian cohort, which comprised workers from the facilities of Atomic Energy of Canada Limited (AECL), were compatible with those for the other national cohorts. In 2005, IARC completed a further study, involving nuclear workers from 15 countries, including Canada. In these results, the dose-related risk for leukaemia was not significant but the prominent finding was a statistically significant excess relative risk per sievert (ERR Sv(-1)) for 'all cancers excluding leukaemia'. Surprisingly, the risk ascribed to the Canadian cohort for all cancers excluding leukaemia, driven by the AECL sub-cohort, was significantly higher than the risk estimate for the 15-country cohort as a whole. We have attempted to identify why the results for the AECL cohort were so discrepant and had such a remarkable influence on the 15-country risk estimate. When considering the issues associated with data on the AECL cohorts and their handling, we noted a striking feature: a major change in outcome of studies that involved Canadian nuclear workers occurred concomitantly with the shift to when data from the National Dose Registry (NDR) of Canada were used directly rather than data from records at AECL. We concluded that an important contributor to the considerable upward shift in apparent risk in the 15-country and other Canadian studies that have been based on the NDR probably relates to pre-1971 data and, in particular, the absence from the NDR of the person-years of workers who had zero doses in the calendar years 1956 to 1970. Our recommendation was for there to be a comprehensive evaluation of the risks from radiation in nuclear industry workers in Canada, organisation by organisation, in which some of the anomalies that we have identified might be addressed.