Explaining deceptive language in topical discourse: Integrating transformer models and explainable AI using LIME and SHAP for transparent text-based deception detection

Detecting deception in textual communication is an important challenge in online environments where misleading information and deceptive narratives are increasingly prevalent. This study investigates the integration of fine-tuned transformer models, prompt-based large language models (LLMs), and explainable AI (XAI) techniques for interpretable deception detection. Its contributions include a statistically validated comparison of six BERT architectures, a factorial prompt ablation across five GPT models, and a quantitative evaluation of LIME (Local Interpretable Model-agnostic Explanations) and SHAP (SHapley Additive exPlanations) explanation quality and stability. Experiments were conducted on two domain-specific datasets totaling 500 statements, with 264 related to climate change and 236 related to COVID-19. The results show that transformer models are effective for deception classification even when trained on relatively small datasets, with DeBERTa achieving 88.87% and 87.92% accuracy on topical datasets. LIME and SHAP were applied to analyze word-level attribution patterns in the best-performing model. The two XAI methods exhibited substantial aggregate-level agreement, converging on similar sets of influential words associated with deception across domains, whereas agreement on individual predictions was more limited. These findings align with prior research on recurring patterns in deceptive language and provide supportive evidence that linguistic patterns identified through explainability techniques remain consistent across topics, even under limited data conditions. Overall, the results demonstrate that combining contemporary language models with XAI techniques can provide both accurate predictions and interpretable insights into the linguistic characteristics of deceptive communication.