AI
New framework lets humans guide AI's fact-checking reasoning in real time
Co-FactChecker enables expert fact-checkers to collaborate with AI models by editing the models' internal reasoning steps, outperforming fully automated approaches.
Researchers have developed a new approach to fact-checking that bridges the gap between human expertise and artificial intelligence by letting experts directly steer how AI models reason through claims.
The collaboration problem#
Professional fact-checkers draw on deep domain knowledge and real-world context that AI systems simply don't possess. While large language models (LLMs) and large reasoning models (LRMs) can process information at scale, they operate primarily within the constraints of available evidence and struggle when feedback needs to guide their thinking. Existing systems that attempt human-AI collaboration rely on dialogue-based exchanges, which prove difficult to calibrate and often produce clunky multi-turn interactions.
How Co-FactChecker works#
The new framework, called Co-FactChecker, introduces a different approach: it treats an AI model's internal reasoning—its "thinking trace"—as a shared workspace. Rather than having experts type corrections in natural language for the model to interpret, the system translates expert feedback into trace-edits: precise, targeted modifications to the model's reasoning steps themselves. This sidestepping of dialogue-based feedback allows experts to inject their knowledge directly into where the AI actually reasons.
Performance and human feedback#
Theoretical analysis suggests trace-editing offers advantages over traditional multi-turn dialogue for steering model behavior. In automatic evaluations, Co-FactChecker surpassed both fully autonomous fact-checking systems and existing human-AI collaboration methods. Human testers further validated the improvement, preferring Co-FactChecker over dialogue-based alternatives. Reviewers noted that the framework produced higher-quality reasoning chains, more accurate verdicts, and thinking traces that were both easier to interpret and more actionable for domain experts.
This work points toward a future where fact-checking remains grounded in human judgment while leveraging the speed and scale that AI systems bring to the table.
