The analysis of error data in short-term memory (STM) tasks has frequently led to the conclusion that semantic information is not stored efficiently in STM. Experiment I used a probe recognition memory task requiring the recognition of either identity or synonymity. Conditions between which the opportunity for semantic confusion errors differed resulted in changes in the rate of false recognition. This evidence was viewed as supporting the hypothesis that semantic information may be stored in STM when such storage is a task demand. A second experiment explored the ability to recognize synonymity when storage in STM is not involved. The results showed less than perfect recognition, indicating that synonyms are only approximate instances of semantic identity.