Advanced data detection will be one of the key enablers to achieving the very high areal recording densities of future tape storage systems. Departing from the partial-response maximum-likelihood (PRML)-based read channel design traditionally used in tape systems, this paper describes noise-predictive maximum-likelihood (NPML) detection, which is a technique that has been known for many years in the hard-disk-drive industry but has been introduced for the first time in the tape storage industry in IBM tape drives. In the NPML read channel design, the readback signals are conditioned prior to data detection so that their noise components are statistically decorrelated and reduced in power. This paper describes the basic principles of NPML detection and its application to tape systems in the form of a 16-state detector. It is argued that, because of the inherent variability of the recording channel characteristics in tape drives, fully adaptive NPML detection needs to be realized in order to optimize detection performance. Actual readback waveforms of data recorded on metal particulate as well as barium-ferrite particulate tape media are used to illustrate the error rate performance achieved by 16-state and 32-state NPML detectors. It is shown that, under realistic worst-case channel conditions, a 16-state NPML detector could offer an improvement in error rate after an error-correcting code of approximately two orders of magnitude.