Abstract <p>Social media, in the era of digital information, is a useful tool for communication. Text categorization, image mining, and evaluation of sentiment have accumulated a lot of attention, partially because of the amount of unstructured data on social networking. Many of these standards are derived from online sharing networks, which provide a vibrant social environment and gather images with cyberbullying posts. These clusters provide rich data that is logically helpful for text retrieval and image classification. It is shown that two important deep learning (DL) and artificial intelligence (AI) methodologies are beneficial for evaluating toxicity from social media interactions (SMI). This paper describes the many procedures involved in toxic word analysis, such as preprocessing, evaluating, and displaying social media content. Using DL approaches, such as bidirectional encoder representations from transformers (BERT) to assess toxic terms in social interactions, the text is first categorized. A BERT-based legislative text classification technique is suggested to identify the policy field defined in the text more precisely. The approach first vectorizes the sentence-level attributes of the legislative text using the BERT-trained language model. The resulting feature vector is then fed into the classifier to do the classification. BERT categorizes text into predetermined categories by first extracting characteristics from the text using filters. The second approach properly identifies photos from social networks using AI techniques. A weighted stack on the social network image association network was created using a neural network. The harmful photographs were then filtered using AI by combining three distinct types of social network photos: RGB (red, green, and blue), grayscale, and depth. Three-dimensional conventional neural networks (3D CNNs) were built to gather and filter social media online bullying posts in real-time during a crisis event to categorize cyberbullying images.</p>