Real-Time Verification and Credibility Analysis via Agentic AI-Driven Chatbot Systems

Author Name : Prof. Manisha Vaidya, Mr. Harsh Ambule, Mr. Komal Shahu, Ms. Samruddhi Bhoyar, Mr. Vinay Dandekar, Ms. Prachi Hukare

DOI: https://doi.org/10.56025/IJARESM.140526113

 

---------------------------------------------------------------

 

ABSTRACT

The rapid growth of online news and social media has intensified the spread of fake news, eroding public trust, polarizing societies, and threatening public safety. Fake news increasingly appears in multimodal formats, combines sophisticated language with compelling narratives, and propagates at a scale that renders manual verification infeasible. Deep learning models and Large Language Models (LLMs) have demonstrated strong performance in text classification and misinformation detection [1][5][8], yet many existing systems lack transparency, real-time operation, and integrated source citation for end-users.

 

This paper presents FactLens, a production-ready, multi-agent, LLM-powered fake news verification system designed to emulate professional fact-checking workflows while remaining accessible via a conversational chat interface. The architecture combines five specialized agents—Ingestion, Classification, Verification, Research, and Final Reasoning—coordinated through an orchestrated pipeline that performs language detection, claim extraction, web scraping, credibility scoring, and evidence-backed verdict generation. FactLens integrates NewsAPI-based news retrieval, JSDOM-driven web scraping, and a Gemini 2.0 Flash LLM core to deliver explainable verdicts with source URLs, confidence scores, and risk levels, supporting up to 25 languages and typical end-to-end response times between 15 and 45 seconds.

 

An evaluation design based on standard information retrieval metrics—accuracy, precision, recall, and F1-score—is described, together with qualitative performance observations from representative health and political misinformation scenarios. The paper also analyzes advantages, limitations, and future extensions such as multimodal fake news detection, integration with domain-specific embeddings, explainable AI (XAI) [10][11], and cross-referencing with professional fact-checking databases. The extended design further studies how heterogeneous evidence returned by LLM web search, GNews API, and NewsAPI is normalized, classified, and aggregated to strengthen robustness against adversarial and low-quality news sources.

 

Keywords: Fake news detection; multi-agent systems; Large Language Models; Gemini 2.0 Flash; misinformation; web scraping; credibility scoring; explainable AI; NewsAPI; LLM agents; FactLens.