Show Me the Work: Fact-Checkers' Requirements for Explainable Automated Fact-Checking

Research Background and Issues

What problems or challenges did the authors identify?

1. The widespread use of large language models (LLMs) and generative AI in online media has significantly increased the volume and complexity of misinformation, making it difficult for fact-checkers to keep up.
2. Current automated fact-checking systems lack explainability, preventing fact-checkers from effectively reviewing system outputs.
3. Existing technologies often prioritize model performance over practical application needs, creating a disconnect with fact-checking practices.

Why is this issue important?

• Misinformation poses a threat to social stability, particularly in political, health, and social media contexts.
• After reaching a peak in the number of fact-checking initiatives, resource shortages have led to a decline in projects, further exacerbating the burden on fact-checking efforts.
• Enhancing the explainability of fact-checking tools is critical to addressing the wave of misinformation driven by generative AI.

Research Motivation and Related Work

• The authors note that fact-checking is a complex task requiring expert knowledge, research skills, and appropriate judgment of sources.
• Fact-checkers remain skeptical of existing automated tools, believing these technologies fail to meet practical needs, especially in the absence of reasonable explanations.
• This study aims to gain deeper insights into fact-checkers' needs to drive the development of truly effective automated tools.

Solutions

What methods or solutions did the authors propose?

1. Through semi-structured interviews with 10 professional fact-checkers from five continents, the authors explored fact-checkers' workflows, decision-making processes, and explanation needs.
2. They provided detailed descriptions and requirements for each step of the fact-checking process, such as selecting claims to verify, retrieving evidence, determining veracity, and disseminating results.

What are the innovative aspects of this solution?

• Direct collaboration with fact-checkers to identify key explanation needs, avoiding the disconnect between researchers' assumptions and real user requirements often seen in traditional studies.
• Introduction of multi-stage explanation needs, including uncertainty, evidence quality, reproducibility, and violation processes.
• Investigation of how fact-checkers use multiple tools to compensate for the limitations of individual tools, suggesting directions for tool integration.

What are the implementation steps? What key technologies were used?

1. Recruitment of fact-checkers from diverse regions and backgrounds, focusing on their perspectives on AI tools and practical usage experiences.
2. Thematic coding and data analysis to summarize fact-checkers' specific needs for explanations and tool performance during their workflows.
3. Integration of existing literature on automated fact-checking to provide concrete recommendations for developing explainable tools.

Research Outcomes

What specific results were achieved?

1. The authors identified four key steps in the fact-checking process: claim selection, evidence retrieval, veracity determination, and result dissemination.
2. Proposed explanation needs include:
   • Providing localized explanations, pointing to specific evidence and indicating sources of uncertainty.
   • Allowing users to verify the reasoning process behind automated tool outputs.
   • Offering relevant background information, explaining the origins and quality of model training data.

How does it compare to existing solutions? What are its advantages?

• Integrates a human-machine collaboration perspective, aiming to complement fact-checkers' work rather than completely replacing human decision-making.
• Provides a detailed breakdown of explanation needs rather than focusing solely on a single stage of the task.
• Emphasizes designing tool logic based on fact-checkers' workflows rather than relying solely on autonomous AI logic.

What were the experimental or evaluation results?

• Fact-checkers' acceptance of AI tools is influenced by the transparency of tool explanations and their alignment with human processes.
• Fact-checkers believe tools can help alleviate workload but note that existing tools lack clarity in explanations and fail to establish trust.
• Current tools perform poorly in addressing the needs of non-English language fact-checking, neglecting linguistic and cultural diversity.

Limitations and Future Directions

Limitations:

• The sample size is relatively small, involving only 10 participants, which may not fully represent global fact-checking practices.
• All interviews were conducted in English, potentially limiting the perspectives of non-native English-speaking fact-checkers.

Future Directions:

1. Explore how tools can be designed to adapt explanations to fact-checking needs across different cultural and linguistic contexts.
2. Investigate methods for quantifying and representing AI tools' "uncertainty" to better align with fact-checkers' intuition and logic.
3. Develop a multi-step technical explanation framework for complex fact-checking processes, covering the entire chain from evidence retrieval to final judgment.
4. Expand global collaboration with fact-checkers to improve technical tools by incorporating multi-regional requirements.

This study provides valuable guidance for the development of automated fact-checking tools, particularly emphasizing the central role of human users in explanation needs.