Several studies have shown that the information conveyed by bell-shaped tuning curves increases as their width decreases, leading to the notion that sharpening of tuning curves improves population codes. This notion, however, is based on assumptions that the noise distribution is independent among neurons and independent of the tuning curve width. Here we reexamine these assumptions in networks of spiking neurons by using orientation selectivity as an example. We compare two principal classes of model: one in which the tuning curves are sharpened through cortical lateral interactions, and one in which they are not. We report that sharpening through lateral interactions does not improve population codes but, on the contrary, leads to a severe loss of information. In addition, the sharpening models generate complicated codes that rely extensively on pairwise correlations. Our study generates several experimental predictions that can be used to distinguish between these two classes of model.