Two of the most popular approaches to induction are instance-based learning (IBL) and rule generation. Their strengths and weaknesses are largely complementary. IBL methods are able to identify small details in the instance space, but have trouble with attributes that are relevant in some parts of the space but not others. Conversely, rule induction methods may overlook small exception regions, but are able to select different attributes in different parts of the instance space. The two methods have been unified in the RISE algorithm (Domingos 1995). RISE views instances as maximally specific rules, forms more general rules by gradually clustering instances of the same class, and classifies a test example by letting the nearest rule win. This approach potentially combines the advantages of rule induction and IBL, and has indeed been observed to be more accurate than each on a large number of benchmark datasets. However, it is important to determine if this performance is indeed due to the hypothesized advantages, and to define the situations in which RISE’s bias will and will not be preferable to those of the individual approaches. This abstract reports experiments to this end in artificial domains.