To utilize DBMSs, a database designer must usually construct a schema, which is used to validate the data stored and help set up efficient access structures. Because database design is an art, and because the real world is irregular, unpredictable, and evolves, truly useful database systems must be tolerant of occasional deviations from the constraints imposed by the schema, including the semantic integrity constraints. We therefore examine the problems involved in accommodating ezceptional information in a database, and ’ outline techniques for resolving them. Furthermore, we consider ways in which the schema can be refined to better characterize reality as it is reflected in the data encountered, including the exceptions. For this purpose, we describe part of a “database administrator’s assistant” a computer system which can suggest modifications and additions to the current definitions and integrity constraints in the schema. This system makes generalizations from the currently encountered exceptions, and is based on techniques used in Machine Learning.