The rapid proliferation of digital communication platforms has created an information ecosystem in which false and misleading content can propagate with greater speed and reach than factually accurate information, with documented consequences for public health, democratic integrity, and social cohesion. The majority of deployed automated detection systems rely on a single analytical signal – typically the linguistic content of a claim – a structural limitation that renders them brittle against claims that present no linguistic anomaly, against unverifiable source credibility, and against “cheapfake” content in which authentic media is presented out of context. This paper presents FactonAI, a real-time misinformation detection system built on a ten-stage forensic verification framework that addresses this limitation through evidential triangulation: the concurrent application of ten analytically distinct verification signals to every submitted claim. The architecture integrates a prompt-injection security gate, fact-check database cross-referencing, open-web evidence mining, multi-modal image forensics, knowledge-graph entity verification, dual-model AI consensus (Google Gemini and HuggingFace-hosted Qwen2.5-7B-Instruct), source credibility scoring, social virality estimation, semantic and psychological manipulation detection, historical archive tracing, and structured verdict synthesis into a single asynchronous pipeline. The resulting verdict is a structured document comprising a classification label, a confidence score, a four-dimensional confidence breakdown, a claim classification taxonomy with harm assessment, a structured evidence base, and a transparent reasoning trace. The system is implemented using Python/FastAPI for the backend and React 19 for the frontend, and is deployed on cloud infrastructure using exclusively perpetual free-tier API services, demonstrating the practical feasibility of comprehensive multi-signal forensic verification without institutional infrastructure investment. A pilot evaluation on 20 pre-registered, ground-truth-labelled claims using the system’s evidence-retrieval-and-reasoning methodology achieved 100% classification accuracy; this result is explicitly scoped to unambiguous, well-documented claims and is reported alongside its limitations rather than generalised, with full adversarial-dataset benchmarking against the deployed dual-model system identified as the principal direction for future empirical validation.
Jaiswal et al. (Wed,) studied this question.