A novel, information-theoretic fuzzy approach to discovering unreliable data in a relational database is presented. A multilevel information-theoretic connectionist network is constructed to evaluate activation functions of partially reliable database values. The degree of value reliability is defined as a fuzzy measure of difference between the maximum attribute activation and the actual value activation. Unreliable values can be removed from the database or corrected to the values predicted by the network. The method is applied to a real-world relational database which is extended to a fuzzy relational database by adding fuzzy attributes representing reliability degrees of crisp attributes. The highest connection weights in the network are translated into meaningful if, then rules. This work aims at improving reliability of data in a relational database by developing a framework for discovering, accessing and correcting lowly reliable data.