Enhancing Disinformation Detection with Explainable AI and Named Entity Replacement

AI systems can be misled by irrelevant details: The research found that disinformation detection models were making decisions based on "superfluous features" like URLs and emoticons rather than actual content quality. This highlights how AI systems can learn shortcuts that don't reflect genuine disinformation patterns, potentially failing when these surface-level cues change.

Explainability tools reveal hidden biases: By using SHAP (a method that explains AI decisions), researchers discovered that models were relying on specific named entities like place names to classify disinformation. This demonstrates why transparency tools are essential for identifying when AI systems develop problematic biases that aren't obvious from performance metrics alone.

Removing entity names improves fairness and generalization: When researchers replaced specific named entities with generic placeholders before training the model, performance on new external data improved by 65.78%. This suggests that AI systems trained on disinformation can unfairly associate certain countries, people, or organizations with false information, and that addressing this bias makes models more reliable across different contexts.

Better preprocessing matters for responsible AI: The study shows that careful data preparation—removing non-informative elements and anonymizing entities—is crucial for building disinformation detection systems that work fairly and reliably. This underscores that responsible AI requires attention to data quality, not just algorithm sophistication.