Explainable AI

Explainable AI (XAI) refers to methods for making AI systems' decisions transparent and understandable to humans. In the context of fake news detection, explainability addresses a critical need: users and fact-checkers must understand why a system flagged content as misinformation to trust the system and potentially correct it.

Explainability approaches operate at different levels:

• Local explanations: why a particular prediction was made for a specific input (e.g., which words or features drove a fake news classification)
• Global explanations: what patterns the model generally uses to make decisions across all inputs
• Rule-based explanations: decompositional or symbolic representations (decision trees, logic rules) that are human-readable by design
• Post-hoc explanations: applying interpretability techniques (LIME, SHAP, attention weights) to already-trained "black-box" models

Key papers

• TELLER: A Trustworthy Framework For Explainable, Generalizable and Controllable Fake News Detection — dual-system approach to explainability: cognition system decomposes claims into interpretable yes/no questions (driven by LLMs), decision system learns transparent logic rules via disjunctive normal form; enables human understanding of reasoning chain and human intervention via rule adjustment
• [[2019-cui-claim-verification-explanations|Cui et al. (2019)]] — provides evidence and explanations for claim verification; uses feature importance and evidence-based reasoning
• [[2023-liao-interpretable-misinformation]] — interpretable multimodal misinformation detection with logic reasoning

Connections

• Trustworthy AI — explainability is one pillar of trustworthiness; others include robustness, fairness, and controllability
• Neural-symbolic AI — neural-symbolic systems are naturally more interpretable than pure neural approaches
• Fake news detection methods — many detection methods lack explainability; new work prioritizes transparent architectures