Automated fake-news detectors have become increasingly accurate, but a new paper on arXiv argues that many of these systems are still black boxes. The authors point out that the models are only weakly connected to established theories of persuasion, credibility, and human information processing, which limits their usefulness in real-world settings where explainability matters.
The paper suggests that disinformation research could benefit from a more deliberate interdisciplinary approach. By harnessing insights from fields such as psychology, sociology, and communication studies, computational models could be designed not only to detect false content but also to explain why it is persuasive. This would make detection systems more transparent and potentially more robust against evolving tactics.
The authors do not propose a single new model or dataset. Instead, they offer a conceptual framework for how to bridge the gap between computational detection and theoretical understanding. The paper is positioned as a call to action for researchers to ground future systems in established knowledge about how people evaluate information and why they believe certain claims.
While the paper does not present experimental results, its argument is timely given the growing demand for accountable AI in content moderation. The authors emphasize that accuracy alone is insufficient if systems cannot justify their decisions to users, regulators, or platform operators. Connecting detection to theory, they argue, is a necessary step toward more trustworthy and interpretable tools.```json // Wait, the user asked for strict JSON with exactly those keys. I need to output valid JSON. I'll put the JSON object directly. No extra text. Let me ensure the body has 2-4 paragraphs. I have 4 paragraphs. Good. Headline under 80 chars: