Learning to Generate and Evaluate Fact-checking Explanations with Transformers

AI can now explain its fact-checking decisions: The research demonstrates that transformer-based AI models can generate human-readable explanations for why they classify information as true or false, moving beyond simple verdicts to transparent reasoning that users can understand and evaluate.

Explanations need quality checks too: The researchers developed automatic evaluation methods to assess AI-generated explanations across multiple dimensions including whether they contradict themselves, contain false information (hallucination), and how convincing they are—recognizing that a fact-check is only useful if its explanation is reliable.

Human judgment remains the gold standard: The study emphasizes the importance of aligning AI-generated explanations with human judgments, suggesting that AI systems should be designed with human evaluation in mind rather than operating independently, which is critical for building user trust in fact-checking tools.

Evidence quality matters significantly: The best-performing model achieved a ROUGE-1 score of 47.77 when provided with high-quality evidence, indicating that fact-checking explanations are substantially more effective when the AI has access to reliable source material—highlighting the importance of evidence curation in misinformation detection systems.

Automation could reduce manual fact-checking burden: By developing metric learning models to automatically evaluate explanation quality, the research suggests a path toward reducing the need for extensive manual assessment of fact-checks, potentially making fact-checking systems more scalable and efficient.