Diazepam (10 mg) was administered to two men performing a delayed matching-to-sample task in which the number of elements in a compound sample stimulus (one of which appeared among 4 comparison stimuli) was manipulated from 1 to 3. After each trial, subjects pressed either a "Yes" or "No" button in response to a computer-presented query about whether the last choice met a point contingency requiring selection of the matching comparison stimulus within a time limit. Diazepam simultaneously produced marginal decreases in matching-to-sample performance and more pronounced decreases in the accuracy of self-reports about the same performance. Diazepam selectively increased false reports of success; false reports of failure were not systematically affected. A signal-detection analysis summarized these patterns as a decrease in self-report discriminability (A') with no systematic change in bias (B'(H)). These preliminary results converge with those of clinical lore and the results of studies with other benzodiazepine drugs in suggesting that diazepam can produce an "overconfidence" in performance self-evaluation, the mechanisms and parameters of which remain to be identified. The results were inconsistent with those of one previous study of diazepam's effects on performance self-evaluation, but given procedural differences between the two studies, the discrepancy may reflect the functional independence of verbal operant classes in Skinner's (1957) taxonomy.