In applied geophysics generalized bodies are often used to model the distribution of underground masses, such as spheres, vertical cylinders, vertical prisms, etc. Generally, discrimination between disturbing bodies producing similar kinds of anomalies is extremely difficult, or even impossible, with classical algorithms. In this paper we present a technique for gravity interpretation, based on the application of feed-forward, multi-layer artificial neural networks (ANNs), trained with back-propagation algorithms. This technique is used, firstly, to discriminate bodies producing a similar kind of anomaly. When the general shape of the body has been found (qualitative interpretation), the ANN method is applied to find the shape parameters like depth, vertical extension and radius. It is shown that after having been properly trained, an ANN is able to recognize a disturbing body with a degree of confidence higher than 99%. It is also shown that inversions carried out with this method produce quantitative results with accuracy ranging from 2% to 5%. The applications presented in this paper are based on synthetic data. These are the first steps towards a generalized technique of interpretation.