A Multi-Stage AI Pipeline for the Moderation and Identification of Harmful Synthetic Content

The fast growth of user-created material on social media sites and the rapid growth of generative AI technology have made it necessary to have strong and scalable automated content moderation systems. Current single-stage moderation methods have high false positive rates and can't tell the difference between real policy infractions and content that has been intentionally created. This is an important distinction that has legal and editorial consequences. This paper introduces a multi-stage AI moderation pipeline aimed at concurrently tackling both difficulties through the integration of semantic pre-filtering, convolutional neural network-based guideline violation detection, and AI-generated picture identification inside a cohesive decision aggregation framework. The first step uses human presence recognition to direct computing resources to images that are sensitive to policy. A fine-tuned MobileNetV2 classifier trained on a curated NSFW dataset finds guideline breaches. At the same time, a second MobileNetV2 model trained on the GenImage benchmark checks the legitimacy of candidate photos. A decision aggregation module combines the results from both branches to make one of three moderation decisions that can be acted on: approve, review, or block. Qualitative evaluation shows that the pipeline works very well in both branches. The AI-detection part got an F1 score of 91.3% on test data that was not used in the training process. The proposed system provides a functional and scalable framework for real-world content moderation operations, influencing platform safety engineering and digital media governance.