Observations of systematic temporal variations of seismic anisotropy are presented for an induced-seismicity experiment at 9 km depth. These observations were made under particularly well-controlled conditions in the German Continental Deep Drilling Program (KTB) borehole, using shear wave splitting from similar events recorded at a three-component instrument located at 4 km depth from a hydraulic fracturing experiment at 9 km depth. In a large set of seismic events recorded during the experiment, many can be associated with multiplets exhibiting essentially identical waveforms. Since they must have approximately the same source location and source radiation pattern, these events are particularly useful for testing the hypothesis of time-dependent anisotropy. Anisotropy itself is clearly a very prominent feature in the data. A simple approach for waveform matching of split shear waves allows unprecedented resolution of variations in shear wave splitting. Importantly, the variation of shear wave splitting with time is a relative measurement, which can be performed with higher accuracy than the associated absolute measurement. In particular, the relative measurement is not affected by timing errors nor by event distance variations. During the experiment the difference between shear wave velocities decreases by ~2% within ~12 hours. After that, the medium apparently approaches a state which is stable for at least 5 hours. We suggest that the temporal variation is due to the tectonic stress release from seismic events caused by the fluid injection. This model requires the presence of fluid-filled cracks at depths larger than 4 km. ¿ 2000 American Geophysical Union