Autism is a neurodevelopmental disorder characterized by difficulties in communication, behavior and social relationships. The earlier it is diagnosed, the greater the chance of a personalized and effective intervention. However, early diagnosis remains a challenge, especially in very young children. In this context, electroencephalography (EEG) emerges as a promising method, as it is non-invasive, relatively affordable and able to reflect neuronal functioning in real time. However, EEG data are complex and difficult for specialists to interpret, which is why AI (artificial intelligence) and machine learning have started to be used more and more frequently. A hybrid CNN+ResNet+BiLSTM deep network was used to classify autistic and normal individuals, and promising detection accuracy was achieved. The problem arises when these models are working as "black boxes" - they provide predictions but how this prediction was caried out is not clear to the user. The paper also explores the application of Explainable AI (XAI) methods, particularly SHAP (SHapley Additive exPlanations), to provide insights into the decision-making process of the AI model. In this study, we used a EEG data and we compare brain waves of two normal and autism groups, including delta, theta, alpha, beta and gamma waves, for an Autism screening test.