Three event-related potential (ERP) experiments examined whether semantic content can be accessed from visually presented words that cannot be consciously identified. Category labels were shown to participants, followed by masked, briefly exposed words that were either exemplars of the category or not exemplars. The task was to verify the category, by guessing if necessary, and to identify the word, naming it if possible. Exposure durations were selected to allow identification in approximately half the trials. For identified words, there was a marked difference in the ERP response between in-category and out-of category words because of an N400 component. For unidentified words, there was a similar although smaller difference. Conscious identification was defined using a variety of approaches: verbal report, 6-alternative forced choice, and binary categorization (in the context of the regression method; A. G. Greenwald, M. R. Klinger, & E. S. Schuh, 1995). By any definition, ERPs for unidentified words showed evidence of semantic processing. In addition, there were differences in the neuronal populations recruited to process above-threshold versus below-threshold words, suggesting qualitative differences.