EVALUATING TRUST AND TRANSPARENCY IN GENERATIVE AI MODELS USING EXPLAINABLE AI TECHNIQUES

Abstract

Generative artificial intelligence (AI) models, such as transformer-based architectures and Generative Adversarial Networks (GANs), have achieved remarkable success in image synthesis and natural language generation. However, their black-box nature raises significant concerns about trust, transparency and interpretability, particularly in critical application domains. This paper provides an experimental model to test trust and transparency in generative AI models through Explainable Artificial Intelligence (XAI) technology. The proposed approach integrates generative models with XAI methods including SHAP, LIME, and attention visualization to provide interpretable information into model behavior. Experiments are conducted on benchmark datasets, including CIFAR-10 and publicly available text datasets, to evaluate the effectiveness of explainability techniques. The evaluation is performed using quantitative metrics such as fidelity, interpretability score, and trust index, along with qualitative user-based analysis. Results indicate that incorporating XAI significantly improves model transparency and user trust while maintaining competitive performance. It is also noted in the study that there is a trade-off between the complexity and explainability of models. The contribution of this work is a practical framework of creating transparent and trustworthy generative AI systems to enable their reliable application in the real world.