The hypothesis that Alzheimer's disease (AD) degrades semantic representations predicts that AD qualitatively alters spontaneous thoughts. In two experiments contrasting free associations to words with strong (e.g., "bride-groom") versus weak (e.g., "body-leg") associates participants with AD produced less common responses (e.g., "bride-pretty") than normal controls but only for words with strong associations, and only on the first (but not on second or third) association response. Furthermore, all participants produced fewer semantically related responses to words with weak associates. Because strong associations should be retrieved more easily than weak associations these results are problematic for retrieval-based accounts of AD. Instead we propose that AD entails a semantic deficit, and that strong associations involve more semantic processing than weak associations (in all speakers).