Dungeons & Deepfakes: Using scenario-based role-play to study journalists' behavior towards using AI-based verification tools for video content

Title of the Paper

Dungeons & Deepfakes: Using scenario-based role-play to study journalists’ behavior towards using AI-based verification tools for video content

Bibliographic Information

• Subject Area: Deepfake videos/media verification technology and journalist behavior research
• Keywords: deepfake, journalism, media verification, role-play, AI tools, human factors, information authenticity

Research Background and Issues

Identified Problems or Challenges

• With the advancement of deepfake technology, manipulative video content has made information verification a significant challenge in the journalism industry.
• Existing deepfake detection tools have limited accuracy, especially in non-ideal conditions, which can lead to fake content being mistaken for real or authentic content being misjudged as fake.
• The operational mechanism of deepfake detection tools mimics antivirus software, providing a "true" or "false" conclusion without detailed explanations, which does not meet the needs of journalistic verification.

Importance of the Research

• Journalists’ behavior significantly impacts the spread of fake content and its influence on the public, making it crucial to study how deepfake detection tools affect journalists' verification behavior.
• The ability to quickly assess the authenticity of video content under the pressure of news deadlines is critical, especially given the significant societal and political implications of deepfake-related events.

Motivation and Related Work

• The increasing accessibility and technological advancement of deepfake tools have made addressing verification challenges in journalism a focal point.
• Research Gap: There is a lack of in-depth studies on journalists’ perceptions of these AI tools and their impact on workflows.
• Motivation: To explore the application and limitations of deepfake tools while considering journalists' behavior and the complexities of the news environment.

Solution

Proposed Method and Innovations:

• Scenario-Based Role-Playing: Inspired by the core gameplay of the tabletop role-playing game Dungeons & Dragons, a flexible semi-structured simulation environment was created. Journalists were invited to participate in news event verification, simulating their real-world workflows and decision-making behaviors.
• Dynamic and Flexible: Unlike traditional structured questionnaires, this method allowed journalists the freedom to choose verification steps, enhancing the authenticity of the study.
• Complex Experimental Design: Immediate-trigger emergency events were incorporated to simulate the high-pressure environment of real-world news scenarios.

Implementation Steps and Techniques:

1. Participants: 24 experienced U.S. journalists with diverse reporting backgrounds (e.g., technology, politics, health).
2. Scenario Setup:
   • Seven virtual scenarios based on real-world news contexts were constructed, such as international political disputes and election crises.
   • Virtual news sources and content were provided, including descriptions of deepfake videos with detection tool results.
3. Demonstration of Verification Tools:
   • Existing deepfake detection tool interfaces (e.g., DeFake) were redesigned to simulate a more realistic user interaction experience.
4. Behavioral Observation:
   • Journalists were encouraged to "think aloud" during verification, recording their behaviors, decision-making patterns, and trust in the tools.
5. Dynamic Adjustments and Urgency:
   • Dynamic news events (e.g., videos going viral suddenly) were introduced to simulate the urgency of news publication.

Research Findings

Specific Outcomes

• Journalists’ Perceptions of AI Tools:
  • Most journalists held a positive attitude toward deepfake detection tools and were willing to incorporate them into their verification toolbox, but they desired improved explainability of the results.
  • The credibility of tool developers (e.g., whether developed by universities or academic teams) played a key role in journalists' trust.
• Characteristics of Verification Workflows:
  • Journalists tended to prioritize traditional verification methods to establish context, resorting to deepfake detection tools only when facing difficulties.
  • For events with high social or political impact, journalists preferred adding more verification steps to avoid errors.
• Risks of Over-Reliance:
  • Some journalists, after initially experiencing high accuracy, became overly reliant on the tools, using them as the final decision-making authority (potentially leading to automation bias and confirmation bias).
  • At times, the results of deepfake detection tools conflicted with other verification methods, causing hesitation in news publication.

Experimental and Evaluation Results

• The role-playing method successfully simulated a highly realistic news verification environment, eliciting natural behaviors from participants.
• Group discussions within the scenario fostered more comprehensive decision-making but were occasionally influenced by the judgments of team leaders.
• The uncertainty introduced by deepfakes led some journalists to prefer publishing "neutral" reports, highlighting incomplete verification rather than making immediate judgments on video authenticity.

Limitations and Future Directions

1. Technical Limitations:
   • Real deepfake videos were not directly used (only textual descriptions), potentially missing opportunities to observe journalists’ intuitive verification abilities.
   • The study was not conducted across regions, possibly failing to fully reflect the diversity of international journalism ecosystems.
2. Future Directions:
   • Explore differences in verification behavior among journalists in various media cultures.
   • Develop comprehensive and explainable deepfake detection tools to help journalists better understand and interpret system results.
   • Expand experimental designs to enable broader applicability of dynamic scenarios through automation.

Conclusion and Significance

• This study emphasizes the need to avoid premature deployment of deepfake detection tools to mitigate risks associated with automation bias.
• The research provides important insights into the application of AI-assisted tools in journalism: enhancing tool explainability and offering training for journalist teams.
• The role-playing methodology demonstrates potential for use in training and behavioral research, with applications extending beyond deepfakes to other AI-assisted work scenarios.