<p>The rapid expansion of digital communication has intensified the challenge of detecting misinformation and spam with reliable confidence. This study presents TRUE-Net, a generative framework that integrates dual-generator blending and Bayesian trust estimation to enhance reliability and interpretability in information authenticity detection. The first generator, based on a variational autoencoder, reconstructs probabilistic features, while the second adversarial generator synthesizes deceptive patterns; their outputs are fused through uncertainty-weighted blending and evaluated by an adaptive trust-scoring discriminator. Experiments on a balanced Twitter dataset demonstrate that TRUE-Net achieves higher accuracy and AUROC than conventional GAN and VAE baselines while producing calibrated confidence scores that reflect prediction reliability. These findings confirm the effectiveness of embedding probabilistic reasoning into generative models and position TRUE-Net as a scalable foundation for developing trustworthy AI systems in social media and misinformation analysis. </p>