The precision and accuracy of trends and seasonal cycles of CO2, as determined from grab samples, was investigated. First, the statistical aspects of infrequent (weekly) sampling were studied by simulating, via a partially random procedure, parallel time series of CO2 flask samples. These simulated flask series were compared to the continuous analyzer records from which they had been derived. The second approach to studying the uncertainties of flask records was to compare real flask results with simultaneous hourly mean concentrations of the in situ analyzers at the Geophysical Monitoring for Climatic Change observatories at Point Barrow, Mauna Loa, Samoa, and the south pole. The latter comparisons emphasized experimental, rather than statistical, errors. The uncertainties and sampling biases depend on the site and on the period of averaging. For monthly means the uncertainty varies from 0.2 to 0.6 ppm (one standard deviation, parts per million by volume), being largest for Barrow. Sampling biases for monthly means at Barrow and Mauna Loa are significant, up to 0.5 ppm. Experimental errors are the dominant error source for annual averages, and spurious interannual variations can be up to 0.4 ppm.