Co-FactChecker: A New Paradigm for Human-AI Collaborative Fact-Checking

This article introduces the Co-FactChecker framework, which uses model thought traces as a shared scratchpad and converts expert feedback into trace edits to enable more efficient human-AI collaborative fact-checking, significantly outperforming traditional conversational interactions. The framework addresses issues in existing methods such as context inflation, ambiguous feedback, and difficulty in locating reasoning steps, combining the model's processing capabilities with expert judgment to improve the accuracy and efficiency of fact-checking.

In the era of information explosion, false information spreads rapidly, professional fact-checking is time-consuming and labor-intensive, and LLMs/LRMs lack real-world grounding. Traditional conversational human-AI collaboration has three major problems: inflated dialogue history increases cognitive load, natural language feedback is ambiguous, and it is difficult to accurately locate reasoning steps that need modification—these issues have spurred the research on Co-FactChecker.

The core innovation of Co-FactChecker is using model thought traces as a shared scratchpad, explicitly unfolding the reasoning process (each reasoning step, hypothesis, and evidence citation is visualized). It introduces a trace editing mechanism: experts can directly insert/delete/modify steps in the thought traces, which is precise and unambiguous. Theoretically, trace editing is more efficient than conversational information transmission, reducing noise and ambiguity while maintaining the integrity and traceability of the reasoning history.

Automatic evaluations show that this framework outperforms fully automated and traditional human-AI collaboration methods on multiple benchmark datasets: the reasoning chain is more rigorous, evidence citations are more accurate, the consistency between judgments and manual annotations is higher, and the number of interaction rounds is fewer. In manual evaluations, experts prefer this framework, considering it highly interpretable, with strong control, and better efficiency; the generated thought traces can also be used as training or audit materials.

Application scenarios include news fact-checking (rapid verification of breaking news), academic research (validation of claims in interdisciplinary literature), corporate compliance (review of marketing/legal documents), and educational training (cultivating critical thinking). Implications for AI design: interpretability is the foundation of collaboration, structured feedback is better than natural language feedback, and human-AI collaboration is superior to full automation.

Current limitations: it focuses on text-based claims and does not cover multimedia content; trace editing has a learning curve for non-professional users; it only handles single claims and does not involve connected argument networks. Future directions: expand to multimodal scenarios, optimize user-friendly interfaces, and handle complex argument networks.

Co-FactChecker overcomes the limitations of traditional conversational interactions and represents an important advancement in human-AI collaborative fact-checking. In today's era of rampant false information, its technical solution combines human judgment with AI capabilities to contribute to building a trustworthy information environment. With the progress of large models and interaction technologies, this system will play a more important role in the future.