Verifiable by construction: evidence-anchored LLMs for explainable fake news detection

The proliferation of sophisticated misinformation threatens societal trust, yet most AI detectors operate as opaque’black boxes,’ lacking the verifiable reasoning essential for human oversight and adoption in high-stakes domains. This critical transparency gap demands a new paradigm where interpretability is a core design principle, not a post-hoc feature. In this work, we propose the Explainable Fake News Detection (XFND) framework, a human-centered pipeline that marries expert-guided feature space validation with evidence-anchored explanation synthesis using large language models. Our approach demonstrably improves feature space separability before training, increasing the silhouette score on public datasets like LIAR by up to 63% (from 0.19 to 0.31). On established benchmarks, the resulting system achieves competitive classification performance, reaching a macro-F1-Score of 0.792 on a binary version of LIAR and 0.731 on PolitiFact, while ensuring outputs are well-calibrated and auditable. We conclude that proactively designing for interpretability enables systems that are both highly accurate and trustworthy by design, establishing a new standard for collaborative AI in the fight against disinformation.