Several thousand three component seismograms from local earthquakes recorded during two field experiments in August 1991 and November 1992 in the Gulf of Corinth have been analyzed to detect shear-wave splitting. After a first selection of the events located in the S window of the considered stations, a second very strict selection of the records is applied in order to avoid the effect of scattered or converted phases which can mimic the behavior of shear wave splitting. Two main directions of fast S-wave polarization have been detected: one oriented N105$^{\circ}$E-N120$^{\circ}$E, the other N55$^{\circ}$E-N75$^{\circ}$E. The first one is perpendicular to the main direction of extension of the Gulf provided by focal mechanism, GPS measurements, and tectonic studies, and is thus consistent with the extensive-dilatancy anisotropy (EDA) model. The second direction is sub-parallel to the direction of the active normal fault closest to the sites. This suggests a local control of the anisotropy by these active faults, either by a local rotation of the total stress field, in which case the EDA model may still explain the anisotropy, or by the existence of a specific micro- or macro-structure generated by the long term fault activity (set of secondary fault planes parallel to the major one ), in which case the anisotropy direction would be significantly rotated from the stress direction (about $50^{\circ}$). The anisotropic signature does not seem to be affected by the geology of the site (pre-Tertiary limestone and Pleistocene sediments), except for a station located on the thick Plio-Quaternary deposits of a delta, where the time delay is significantly larger.