The proliferation of misinformation poses significant challenges to public discourse and decision-making. While human fact-checkers provide high-quality verification, they cannot scale to match the volume of claims circulating online.
Automated fact-checking systems are often regarded as a promising solution. However, existing systems face a tradeoff: broad evidence retrieval risks including unreliable sources, while narrow retrieval misses critical context. Many approaches either apply simple RAG in one shot, decompose claims but lose contextual dependencies, or apply source credibility signals monolithically without adapting to claim structure.
How can we design an agentic system that balances evidence comprehensiveness with source reliability for verifying complex, real-world claims?
Solution
We built ClaimCLAIRE, an agentic AI system to fact-check claims, analyze source reliability, and provide trustworthy fact-checking results. Here is our quick demo!
In short, trust-modulated retrieval provides the largest gain. Decomposition is insufficient alone but essential when combined with trust-aware retrieval and adaptive gap-filling.
Our error analysis also revealed that incorrect predictions are largely attributed to insufficient claim context, circular reporting, semantic misalignment, and web search API fluctuations.
