Although simulated patients have been increasingly used in medical education, the validity of this procedure--that is, the relationship between performance on a simulated patient and performance with a real patient presenting a similar clinical problem--has not been adequately addressed. The study on which this article is based was an attempt to examine this question. Four actual patients with chronic stable findings were used in the study; four simulated patients were then programmed to present the same problem. A sample of 10 residents in family and internal medicine interviewed and examined all eight patients. The order of the presentation of the cases was balanced. No significant differences emerged in number of questions on history and physical examination or in the diagnoses and investigations considered by residents. Residents elicited significantly more historical data from the simulated patient; however, this was found to be due to a single case in which the real patient suffered from a neurological condition characterized by loss of memory. Residents correctly identified 67 percent of the patients as real or simulated against a chance figure of 50 percent. The implications of the study for training of simulated patients are discussed.