The advantages of magnetic gradiometry as an adjunct to total field mapping are generally recognized and a few aircraft have been equipped with gradiometers. These gradiometers are derived from high-sensitivity total-field magnetometer systems that are in themselves subject to certain errors that can usually be tolerated in conventional surveys. However, in a gradiometer, where very large total-field values are differenced, these errors can, in many cases, greatly exceed the basic accuracy required of the system.There are two principal sources of error in inboard gradiometer systems. The first, and most significant, results from the inevitable magnetic interference of the aircraft or from the inability of currently available compensation systems to deal with the magnetic interference adequately. Passive methods of compensation are not sufficiently comprehensive for gradiometry and the active compensation systems currently in use, which were designed for military applications, cannot guarantee compensation at zero frequency (dc) or at the very low frequencies of interest to the geophysicist concerned with long-wavelength anomalies. The second source of error is the frequency-counting technique usually employed to convert a Larmor frequency to ambient total field. The counting process is somewhat analogous to digital sampling at a relatively low rate and as such affords little protection against aliasing from higher frequency interference sources, including components at aircraft maneuvering frequencies.This paper, using examples, illustrates the two types of error. A list of design criteria is presented and several techniques are described for realizing these criteria. Finally, compensation and survey line results are shown for a three-axis gradiometer system in the National Research Council of Canada's Convair 580. This aircraft uses nonoriented cesium magnetometers, one in each wingtip and one at the tip of the tail fin. Compensations over the entire normal maneuver envelope of the aircraft on all headings give typical standard deviation errors of 3 mgamma /m from dc to 1 Hz. Thus, the system is capable of measuring gradients down to nongeologic background levels.