XAI-Grounded Explanation Generation for Speech Deepfake Detection with Training-Free Multimodal Large Language Models

XAI methods significantly improve explanation quality: When explainable AI techniques (like gradient-based attribution) are combined with language models, the accuracy of generated explanations increases by over 45%, making AI decisions more trustworthy and understandable to users.

Natural language explanations are more useful than technical ones: The paper demonstrates that people better understand explanations written in natural language compared to traditional technical attribution signals, highlighting the importance of making AI systems communicate in human-friendly terms rather than raw data visualizations.

Grounded explanations require evidence, not just language: Generic explanations from large language models alone are insufficient—explanations must be anchored in specific, task-relevant evidence (like audio features in deepfake detection) to be truly reliable and useful for decision-making.

Training-free approaches can enhance existing systems: The framework doesn't require retraining AI models, meaning organizations can improve their speech deepfake detection systems' explainability by integrating existing XAI methods with multimodal language models as an add-on layer.

Human evaluation validates AI explanations: The researchers verified their results through human evaluation and faithfulness checks, emphasizing that responsible AI development requires testing whether explanations actually make sense to people, not just whether they satisfy technical metrics.