A method, apparatus and computer program for detecting natural language (NL) bias by a conversational system is described. Embodiments of the invention determine an NL bias in a set of training questions used to train a machine learning model used by the conversational system to select a user intent. Other embodiments of the invention determine an NL bias in a user question received by the conversation system as compared to the set of training questions. The NL bias causes the machine learning to preferentially associate user queries to a user intent. In respective embodiments, the system takes a corrective action to adjust the NL bias of the training questions or the user question.