Seismic data are rich in information about the structure of geological formations and their pore content. Seismic attributes are any information that can be derived or computed from seismic data. Over the past decade, the increasing number of seismic attributes itself becomes a challenge for interpreter and even for machines to accomplish a reliable interpretation. As in oil exploration wrong decisions may result in costly dry wells, a good understanding of the seismic attributes and the limitations of the machine is of great importance. In this paper, we conduct a brief comparison between deterministic approaches by human and statistical approaches established by machine intelligence. Then, a real example from North Western Australia is presented where seismic attributes are used to find hydrocarbon-saturated rocks and discriminate them from their surrounding rocks. Machine learning, namely, artificial neural network (ANN) is used to combine the seismic attributes to produce hydrocarbon probability cube from seismic data. A number of seismic attributes are used in the training of the neural network. Further, we propose a new seismic attribute combining the gradient and the scaled-Poisson reflectivity which are both sensitive to fluid. The new attribute showed a better detection than the neural network. To benefit from the statistical relationship determined by the ANN and from the proposed deterministic attribute we added the attribute to the ANN training. A better detection of the gas-saturated reservoir is achieved. We conclude that that machine learning algorithms can reduce the workload and save computation time but they need assistance from interpreter to make good decisions.