Gas transfer velocities of SF6 and 3He were determined in a Kerguelen Islands lake at wind speeds in the range 0--10 m/s by injecting the two tracers into the water and measuring their concentrations over 40 days. Two methods are investigated for the determination of the relationship linking the gas transfer velocity K to the wind speed W. The first method postulates a power law relationship K=βWα. This leads to the same exponent α=1.5¿0.2 for both gases. The second method is the classic determination of the gas transfer velocity Kij between two tracer measurements at times ti and tj using the well-mixed reservoir assumption. This method proves to be less favorable owing to its nonlinearity bias and also because it induces much scatter in the gas transfer coefficient/wind speed relationship. This dispersion is shown to arise from the experimental scatter of the data and, above all, from the high sensitivity of the method to even small heterogeneities in the tracer vertical distribution. In the present experiment, the Liss and Merlivat correlation <Liss and Merlivat, 1986> is shown to underestimate the actual mean gas exchange rate by about 40%. Our results agree with the recent dual-tracer experiment by Wanninkhof et al. <1993> and are also consistent with CO2 transfer coefficient data derived from the study of 14C oceanic inventories. As expected from gas transfer theories and various experimental observations, the Schmidt number exponent in the comparison of 3He and SF6 transfer velocities is found to vary substantially with the transfer regime. However, its variation is found to be greater than that forecast by current gas transfer models, with values as high as n=-0.2 for intermediate to strong winds. This again raises the question of the validity of the normalization method for KCO2 calculation from gas transfer experiments, especially in high-wind regimes. ¿ 2000 American Geophysical Union