Breaking News or Breaking Trust? Exploring Challenges and a Design Space for Trustworthy LLM Integration in Journalism

Paper Title

Breaking News or Breaking Trust? Exploring Challenges and a Design Space for Trustworthy LLM Integration in Journalism

Publication Info

• Topic area: Trustworthy integration of Large Language Models (LLMs) in journalism.
• Keywords: LLMs, journalism, trustworthiness, AI literacy, participatory design, explainable AI (XAI), fact-checking, neutrality, autonomy, efficiency.

Background and Problem

• Problem / challenge: The integration of LLMs into journalism introduces risks to journalistic values such as factuality, neutrality, autonomy, efficiency, and AI literacy. Existing guidelines for responsible use of LLMs are abstract and difficult to apply in practice.
• Significance: Journalism is a cornerstone of democratic societies, and ensuring trustworthiness in LLM-infused tools is critical to maintaining public trust and ethical standards in news production.
• Motivation and related work: Prior studies have explored generative AI’s impact on journalism, identifying risks like hallucinations, biases, and inefficiencies. However, there is limited research on designing tools to concretize abstract guidelines and support trustworthy LLM integration.

Solution

• Proposed approach: A design space framework addressing five challenges (factuality, neutrality, autonomy, efficiency, AI literacy) and iterative prototyping of interactive tools for trustworthy LLM-infused journalism.
• Novelty:
  1. Identification of five key challenges for trustworthy LLM integration in journalism.
  2. Development of four prototypes (Fact-Checker, Emphasis Manipulator, Uncertainty Visualizer, Attention Visualizer) contextualized to journalistic practices.
  3. Qualitative evaluations of prototypes with diverse news industry stakeholders.
  4. Exploration of how interactive tools can concretize abstract guidelines and embed XAI techniques into journalism tools.
• Procedure and key techniques:
  • Conducted eight semi-structured interviews and two contextual interviews with journalists, educators, and media house developers.
  • Mapped a design space based on identified challenges.
  • Iteratively developed and evaluated four prototypes using participatory design methods.
  • Prototypes leveraged techniques such as named entity recognition, sentiment analysis, token confidence visualization, and attention mechanism simulation.

Results

• Concrete findings:
  • The Fact-Checker was perceived as useful for systematic fact-checking and concretizing human-in-the-loop guidelines.
  • The Uncertainty Visualizer and Attention Visualizer raised awareness of LLMs’ statistical nature but added limited immediate value to journalistic workflows.
  • The Emphasis Manipulator provided an intuitive alternative for journalists with limited prompting skills but was slower for experienced users.
• Advantage over baselines:
  • Prototypes contextualized known techniques (e.g., XAI, direct manipulation) to journalism, addressing specific challenges like hallucinations, bias, and inefficiency.
  • Tools like the Fact-Checker provided structured approaches to fact-checking, reducing cognitive load and error risks.
• Experiments / evaluation:
  • Evaluations included qualitative feedback from journalists, educators, and developers.
  • Tools were assessed for usability, relevance to identified challenges, and alignment with journalistic values.
• Limitations and future work:
  • Limited participant diversity (two media houses, few journalists).
  • Prototypes like the Attention Visualizer were simulations due to API constraints.
  • Future work includes expanding participant scope, refining prototypes, and exploring alternative methods for improving AI literacy.

Summary

This study identifies five challenges for trustworthy LLM integration in journalism—factuality, neutrality, autonomy, efficiency, and AI literacy—and explores them through a design space and four prototypes. The Fact-Checker was particularly effective in concretizing guidelines for fact-checking, while tools like the Uncertainty Visualizer and Attention Visualizer highlighted the statistical nature of LLMs but faced adoption barriers due to time constraints in journalistic workflows. The findings underscore the need for flexible tools tailored to diverse stakeholders and suggest that journalist education may be a better context for improving AI literacy. Future research should expand on this design space and investigate scalable solutions for trustworthy LLM integration in journalism.