Linear least squares estimation techniques can be used to enhance suppression of narrowband interference in direct sequence spread spectrum systems. Nonlinear techniques for this purpose have also been investigated recently. In this paper, we derive maximum likelihood receivers for direct sequence signal in Gaussian interference with known second order characteristics. It is shown that if the receiver uses samples from outside the bit interval, then the receiver structure is nonlinear. The bit error rate performances of these receivers are compared to those of linear receivers employing one-sided and two-sided least squares estimation filters, for the case of Gaussian autoregressive interference. The results in this paper show that intersymbol interference due to filter taps extending beyond the bit interval cannot be ignored for small processing gains. In some cases, not accounting for intersymbol interference yields too optimistic error estimates, very much away from the true error rates.