While automatic spike sorting has been investigated for decades, little attention has been allotted to consistent evaluation criteria that will automatically determine whether a cluster of spikes represents the activity of a single cell or a multiunit. Consequently, the main tool for evaluation has remained visual inspection by a human. This paper quantifies the visual inspection process. The results are well-defined criteria for evaluation, which are mainly based on visual features of the spike waveform, and an automatic adaptive algorithm that learns the classification by a given human and can apply similar visual characteristics for classification of new data. To evaluate the suggested criteria, we recorded the activity of 1652 units (single cells and multiunits) from the cerebrum of 12 human patients undergoing evaluation for epilepsy surgery requiring implantation of chronic intracranial depth electrodes. The proposed method performed similar to human classifiers and obtained significantly higher accuracy than two existing methods (three variants of each). Evaluation on two synthetic datasets is also provided. The criteria are suggested as a standard for evaluation of the quality of separation that will allow comparison between different studies. The proposed algorithm is suitable for real-time operation and as such may allow brain–computer interfaces to treat single cells differently than multiunits.