Synchronization of neural oscillations has been hypothesized to play an essential role in the large-scale integration of the activity of distinct cell assemblies in the brain. Interestingly, fast-frequency oscillatory episodes in vivo are short-lived and lose temporal coherence over the scale of several tenths of milliseconds. This could be the result of noisy fluctuating inputs extrinsic to the cortex, but this would require substantial spatial correlations in these fluctuations on the scale of the local circuits generating the rhythm. In the present modeling work we explore the hypothesis that temporal decoherence is generated intrinsically within the cortex. We study a reduced model of a piece of cortex which consists of two local networks, each of them comprising two populations of neurons, one inhibitory and one excitatory. As a consequence of the longer range of the excitation the two networks interact via their excitatory populations. Oscillations in the gamma frequency band are generated by the local delayed inhibition in each of the networks. We find that the long-range excitation is responsible for the generation of a variety of complex multi-population phase-locked rhythms, whose exact nature depends on the strength of the local inhibition. Remarkably, when the long-range excitation is sufficiently strong, the activity of the whole system displays synchronous chaotic oscillations with the desired fast temporal decorrelation times. Furthermore, a generalization of the basic model can be used to predict dynamical effects in the visual striatal cortex (eg., cat or macaque monkey), and this provides the ground for a possible experimental verification of our theory.