We explore and discuss several important issues concerning the derivation of many-body force fields from ab initio quantum chemical data, In particular, we seek a general methodology for constructing ab initio force fields that are "chemically accurate" and are computationally efficient for large-scale molecular dynamics simulations. We investigate two approaches for modeling many-body interactions in extended molecular systems. The interactions are adjusted to reproduce the many-body energy in small molecular clusters. Subsequently, the potential parameters affecting only pair interactions are then varied to reproduce the ab initio binding energy of dimers. This simple procedure is demonstrated in the design of a new polarizable force field of water. In particular, this new model incorporates the usual many-body interactions due to electrostatic polarization and a type of nonelectrostatic many-body interactions exhibited in bifurcated hydrogen-bonded systems. The static and dynamical properties predicted by the new ab initio water potential are in good agreement with the successful empirical fluctuating-charge potential of Rick et al. and with experiment. The aforementioned "cluster" approach is compared with an alternative method, which regards many-body interactions as manifestations of the electrostatic polarization properties of individual molecules. The effort required to build ab initio databases for force field parametrization is substantially reduced in this alternative method since only the monomer properties are of interest, We found intriguing differences between these two approaches. Finally our results point to the importance of discriminating ab initio data for force field parametrization. This is essentially a consequence of the simple functional forms employed to model molecular interactions, and is inevitable for large-scale molecular dynamics simulations.