This study examined the question whether semantic content can be accessed from words that are displayed under conditions (brief exposure, pattern masking) where they are not consciously identified. Thirty subjects performed a combined categorization and word identification task. Category labels (e.g. "fruit") were shown followed by masked, briefly exposed words, either validly cued by the category (e.g. "apple") or invalidly cued (e.g. "horse"). The subjects' task was to verify the category by pressing a button, and to identify the word by a verbal response. Exposure durations were selected to allow the subjects to identify approximately 50 % of the words. Four ERP averages were formed: for identified/unidentified words * validly/invalidly cued words. For identified words, there was a marked difference in the ERP response between validly and invalidly cued words. The difference was due to an N400 component for the invalidly cued words. For unidentified words, there was also a significant difference between validly and invalidly cued words, although smaller in size (23 % of that for the identified words). It had the same direction, and appeared in the same time range and with the same general topography as for the identified words. The results show that some semantic processing can operate on words that are not consciously identified