SUMMARY Seismic attributes are derived measures from seismic data that help characterize subsurface geological features and enhance the interpretation of subsurface structures: we propose to exploit the hidden layers of Long–Short Time Memory neural network predictions as possible new reflection seismic attributes. The idea is based on the inference process of a neural network, which in its hidden layers stores information related to different features embedded in the input data and which usually are not considered. Neural network applications typically ignore such intermediate steps because the main interest lies in the final output, which is considered as the exclusive exploitable feature of the process. On the contrary, here we analyse the possibility to exploit the intermediate prediction steps, hereafter referred as ‘deep attributes’ because they are produced by a deep learning algorithm, to highlight features and emphasize characteristics embedded in the data but neither recognizable by traditional interpretation, nor evident with classical attributes or multi-attribute approaches. Nowadays, classical signal attributes are numerous and used for different purposes; we here propose an original strategy to calculate attributes previously never exploited, which are potentially complementary or a good alternative to the classical ones. We tested the proposed procedure on synthetic and field 2-D and 3-D reflection seismic data sets to test and demonstrate the stability, affordability and versatility of the entire approach. Furthermore, we evaluated the performance of deep attributes on a 4-D seismic data set to assess the applicability and effectiveness for time-monitoring purposes and comparing them with the sweetness attribute.