Although current AMBER force fields are relatively accurate for canonical B-DNA, many non-canonical structures are still described incorrectly. As non-canonical motifs are attracting increasing attention due to the role they play in living organisms, further improvement is desirable. Here, we have chosen Z-DNA molecule, can be considered a touchstone of the universality of empirical force fields, since the non-canonical α and γ backbone conformations native to Z-DNA are also found in protein-DNA complexes, i-motif DNA and other non-canonical DNAs. We show that spurious α/γ conformations occurring in simulations with current AMBER force fields, OL15 and bsc1, are largely due to inaccurate α/γ parameterization. Moreover, stabilization of native Z-DNA substates involving γ = trans conformations appears to be in conflict with the correct description of the canonical B-DNA structure. Because the balance of the native and spurious conformations is influenced by non-additive effects, this is a difficult case for an additive dihedral energy scheme such as AMBER. We propose new α/γ parameters, denoted OL21, and show that they improve the stability of native α/γ Z-DNA substates while keeping the canonical DNA description virtually unchanged, and thus represent a reasonable compromise within the additive force field framework. Although further extensive testing is needed, the new modification appears to be a promising step towards a more reliable description of non-canonical DNA motifs and provides the best performance for Z-DNA molecules among current AMBER force fields.