Auditory event-related potentials (ERPs) were recorded in a "double oddball" paradigm requiring an easy and a hard pitch discrimination from multiple sclerosis (MS) patients with and without dementia, and a group of age and sex matched normal subjects. Cognitive function was assessed by a short battery of neuropsychologic (NP) tests, and the two groups of MS patients were selected on the basis of substantial non-overlapping degrees of cognitive deficit in the demented as compared to the non-demented group. The N100, P200 and P300 ERP components were longer in latency in the demented patients, and the N100-P300 interval was prolonged as well, compared to the non-demented patients, whose ERP latencies did not differ from those of the normal subjects. Increased P300 latency was associated with poorer performance on the NP tests, especially those sensitive to impairment of learning and retrieval from memory. The reaction times of both patient groups were prolonged as compared to the controls, whereas the accuracy of the demented patients was significantly poorer than that of the non-demented patients.