A Stacked Ensemble Framework for Detecting Generative AI–Driven Misinformation Using Multi-Modal Social Features

Abstract

The rapid speed of generative artificial intelligence (GenAI) development, the creation and distribution of artificial and misinformation on the Internet have increased considerably. The common misinformation detection systems are based on traditional unimodal textual features and single model classifiers, which means that they do not work well when faced with AI-generated or doctored posts. The proposed research work is a stacked ensemble model that will identify Generative AI-motivated misinformation based on a multi-modal system of social, linguistic and credibility-based features. The method is tested on a filtered collection of 500 social media posts labeled with 31 different features; such as the sentiment polarity, toxicity, readability, author metadata, synthetic-content detection scores, fact-checking verdicts and engagement metrics. Though each of the individual base learners as LightGBM, Logistic Regression and several deep neural networks has a moderate performance (AUC 0.55-0.58), their joint prediction can be used as excellent meta-features. A trained LightGBM meta-learner on these results provides a significant performance improvement; with a ROC-AUC of 0.968, accuracy of 92% and F1-score of 0.926, which is better than all other stand-alone models. Explainability via SHAP also serves as a confirmation of the fact that stacked predictions are most useful in final classification. The findings indicate that weak models; in their combination can provide a potent answer to modern GenAI-motivated misinformation and multi-model stacking is crucial in the analysis of misinformation in the future.