The Argos satellite system is commonly used to track and relay behavioral data from marine mammals, but their underwater habit results in a high proportion of locations of non-guaranteed accuracy (location classes (LC) O, A, and B). The accuracy of these locations is poorly documented in marine mammals. We assessed the accuracy of all LCs on four juvenile gray seals fitted with Argos satellite relay data loggers and held in captivity in an outdoor tank for a total of 61 seal-days. Four hundred and twenty-six locations were obtained from seals in captivity, and their latitude and longitude error was assessed before and after filtering, following MConnell et al. (1992). There was significantly more error in longitude than latitude in all LCs except I. C A. The ratio of the standard deviations of longitude : latitude ranged from 1.77 (LC 3) to 2.58 (LC 1). Filtering had very little effect on errors in LCs 3-1, but in the remaining LCs filtering resulted in error reductions ranging from 8% to 63%. In LCs O, A, and B, error reduction was greater in the 95th percentile errors, especially in longitude. The averages of the latitude and longitude 68th percentile errors and those predicted by Argos (in brackets) were 226 (150), 372 (350), and 757 (1000) m for LCs 3, 2, and 1 respectively. Both latitude and longitude errors of LCs > O were normally distributed. Both filtered and unfiltered LC A locations were of a similar accuracy to LC 1 locations, and considerably better than LC O locations.