The preservation of the authenticity of evidence on digital platforms is significantly hampered by the spread of disinformation. This challenge is exacerbated by the complex linguistic patterns and contextual nuances that define misinformation. High generalization and predictive accuracy across a range of datasets are sometimes tricky for traditional deep learning models, revealing significant application limitations. To address these challenges, this paper presents a strong ensemble framework that makes use of leading deep learning architectures. The suggested methodology consists of pre-trained Glove and BERT embedding and includes convolutional neural networks (CNN), multilayer perceptron (MLP), and bidirectional long short-term memory (Bi-LSTM) networks. As these embedding precisely reflect semantic and contextual language representations, they enable a deeper understanding of textual content. Techniques like weighted voting, hard voting, and an optimized logistic regression model enhanced via threshold and hyper parameter optimization are used in integrating the ensemble models. Three easily accessible datasets—PolitiFact, University of Victoria, and Gossip Cop—were used to evaluate the suggested analysis method. The optimized logistic regression model performed better on two datasets, achieving F1 scores of 85.4% on PolitiFact and 99.51% in University of Victoria data, which are consistent with the experimental findings.