MindBugs: Self-Explanatory Disinformation Detection System with Human-in-the-loop

We introduce a self-explanatory, human-in-the-loop system for disinformation detection, that shifts the focus from binary classification to structured narrative understanding. Central to our approach is a hierarchical framework that organizes disinformation into interconnected narrative structures. This design not only enhances transparency, but also enables domain experts to intervene drastically in the classification process. We also present VER-1, a dataset of Eastern European disinformation and war propaganda, and evaluate our system on it alongside a COVID-19 corpus. The system performs best on VER-1, reflecting its focus on structured, campaign-driven disinformation rather than isolated falsehoods. The solution combines large language models with hierarchical clustering and textual entailment to construct and verify narrative structures. At each clustering step, the vector space is refined through synthetic data points, enabling semantically aligned texts to merge. This supports classification by narrative alignment and results in an interpretable, correctable, and expert-guided alternative to black-box models.