The problem of demonstrating that natural language mediators play a role in learning and are not epiphenomena resulting from learning is an important problem in cognitive learning theories. Using a cued-recall and a free-recall learning task, Ss were requested to add a one-word mediator to some of the pairs of concrete nouns presented, The mediated pairs were learned better than the control pairs in both tasks. Both words were recalled only when the mediator was also recalled. Also, one-word mediators were the most effective recall cues and were the best recalled words in free recall. A two-stage learning model adequately described the data. However, a counterargument can be made which considers the mediator to be a high associate of one of the words presented and actually has no direct link to the other presented word. A possible experimental resolution of the problem is discussed.