In a paired associate learning task, subjects responded to each presentation of a nonsense syllable by typing both a three-letter associate and a rating of their confidence that this response was correct or incorrect. Average event-related potentials (ERPs) elicited by the subsequent presentation of the actual paired syllable varied with the interaction of confidence and trial outcome. A larger amplitude P300 was elicited by syllables that informed subjects that they were correct when they thought they were incorrect or that they were incorrect when they thought they were correct than by syllables that confirmed subjects expectations. That this average ERP result was indeed an effect on P300 amplitude, and not an artifact of single-trial variability in P30O latency, was confirmed with a trial-by-trial latency adjustment procedure. Consistent with findings from other tasks, P300 amplitude varied inversely with the subj ective probability of the ERP-eliciting events.