The digital accelerograph network installed in Taiwan produced a rich set of records from the 20 September 1999 Chi-Chi, Taiwan earthquake (Mw 7.6). Teledyne Geotech model A-800 and A-900A* digital accelerographs were colocated at 22 stations that recorded this event. Comparisons of the amplitudes, frequency content, and baseline offsets show that records from several of the A-800 accelerographs are considerably different than those from the colocated A-900A accelerographs. On this basis, and in view of the more thorough predeployment testing of the newer A-900A instruments, we recommend that the records from the A-800 instruments be used with caution in analyses of the mainshock and aftershocks. At the Hualien seismic station two A-900A and one A-800 instruments were colocated, along with a Global Positioning System instrument. Although the records from the two A-900A instruments are much more similar than those from a colocated A-800 instrument, both three-component records contain unpredictable baseline offsets, which produced completely unrealistic ground displacements derived from the accelerations by double integration, as do many of the strong-motion data from this event; the details of the baseline offsets differ considerably on the two three-component records. There are probably numerous sources of the baseline offsets, including sources external to the instruments, such as tilting or rotation of the ground, and sources internal to the instruments, such as electrical or mechanical hysteresis in the sensors. For the two colocated A-900A records at the Hualien seismic station, however, the differences in the baseline offsets suggest that the principal source is some transient disturbance within the instrument. The baseline offsets generally manifest themselves in the acceleration time series as pulses or steps, either singly or in combination. We find a 0.015-Hz low-cut filter can almost completely eliminate the effects of the baseline offsets, but then information regarding the permanent displacements is lost. The causative mechanisms of the baseline offsets are unknown presently. Hence, it is very difficult to recover the permanent displacements from the modern digital records, although for records close to large earthquakes, the signal-to-noise ratio should theoretically be adequate to obtain ground motions with periods of hundreds of seconds. This study reinforces our conclusion from previous studies that the sources of baseline offsets occurring in digital strong-motion records are very complex and often unpredictable, and that, therefore, it is difficult to remove the baseline effects to maximize the information content of the record. The baseline offsets only affect very long period motions (e.g., >20 sec), however, and therefore are of little or no engineering concern.