Dealing with Uncertainty: Understanding the Impact of Prognostic Versus Diagnostic Tasks on Trust and Reliance in Human-AI Decision Making

Title of the Paper

Dealing with Uncertainty: Understanding the Impact of Prognostic Versus Diagnostic Tasks on Trust and Reliance in Human-AI Decision-Making

Paper Information

• Subject Area: Human-Computer Interaction, Trust and Reliance in Artificial Intelligence, Task Characteristics Analysis
• Keywords: AI trust, task complexity, task uncertainty, diagnostic tasks, prognostic tasks, binary decision-making, appropriate reliance, human-AI collaboration, research methodology

Research Background and Issues

• Issues and Challenges:

  • Existing research has made significant progress in exploring factors influencing human trust and reliance on AI systems (e.g., human expertise, AI system characteristics). However, there is a lack of systematic studies on how task characteristics (such as complexity and uncertainty) shape human-AI decision-making.
  • In scenarios with high complexity and uncertainty, individuals experience greater cognitive load, which may lead to inappropriate over-reliance or under-reliance on AI systems.

• Research Importance:

  • Understanding the impact of task complexity and uncertainty on trust and reliance in AI systems can aid in designing AI tools and intervention mechanisms better suited for complex and uncertain scenarios, thereby improving human-AI team collaboration efficiency.

• Research Motivation and Related Work:

  • Diagnostic tasks (complete information and currently verifiable) and prognostic tasks (incomplete information requiring future outcome prediction) introduce different dimensions of task uncertainty.
  • Humans tend to rely more on AI in complex tasks, but task complexity and uncertainty may independently or interactively influence trust and reliance systems.
  • This research aims to explore how to design experiments that control task characteristics to promote appropriate human reliance on AI.

Solution

• Proposed Method:

  • Designed a 3×2 experimental analysis based on a "travel planning" scenario, covering three levels of task complexity (low, medium, high) and two types of task uncertainty (diagnostic tasks and prognostic tasks).
  • Used an AI system with a 65.7% reliability rate to simulate reliance behavior on AI suggestions in a travel route planning context.

• Innovativeness:

  • For the first time, experimentally validated the interaction effects of task complexity and task uncertainty in real-world scenarios.
  • Proposed a framework for modeling task uncertainty through diagnostic and prognostic tasks, contributing to a more scientific evaluation of human-AI interaction.

• Implementation Steps and Techniques:

  • Developed a task complexity framework: defined complexity levels based on the number of task constraints.
  • Developed a task uncertainty framework: distinguished experimental conditions through complete information (diagnostic tasks) and ambiguity introduced by predicting future information (prognostic tasks).
  • Used performance metrics (e.g., accuracy, reliance indicators) to analyze user reliance on AI systems and decision-making performance under different experimental conditions.

Research Findings

• Specific Findings:

  • Task complexity and uncertainty significantly impact human reliance on AI systems.
  • Participants exhibited higher reliance on AI suggestions in high-complexity tasks, but the level of appropriate reliance decreased.
  • In prognostic tasks, uncertainty increased participants' over-reliance on AI systems while weakening their ability to discern correct AI suggestions.
  • A turning point for appropriate reliance emerged in high-complexity and prognostic tasks, where participants relied more appropriately on AI, leading to improved overall accuracy.

• Comparison with Existing Solutions and Advantages:

  • Compared to existing work, this study not only focuses on low-complexity or low-uncertainty tasks but also introduces more complex and realistic task scenarios to validate dynamic changes in appropriate reliance.
  • Introduced a new perspective of diagnostic and prognostic tasks, providing a more refined framework for uncertainty assessment.

• Experimental or Evaluation Results:

  • In high-complexity and high-uncertainty tasks, the turning point for appropriate reliance significantly improved the accuracy of AI-assisted decision-making.
  • Interaction effects were tested to examine the combined influence of complexity and uncertainty, confirming the significant impact of task characteristics on reliance behavior.

• Limitations and Future Directions:

  • Limitations:
    • Task design may be influenced by cognitive biases (e.g., overconfidence or familiarity bias).
    • Experimental conditions did not fully cover all possible scenarios of uncertainty and complexity in the real world (e.g., conflicting information).
  • Future Directions:
    • Extend research to more domains and task types to explore the effects of complexity and uncertainty.
    • Investigate how to effectively present AI's explanatory information to improve user decision quality in complex and uncertain tasks.

Conclusion

This study provides new perspectives and empirical evidence for understanding how task complexity and uncertainty influence trust and reliance. It suggests that researchers should focus on systematic evaluation and modeling of task characteristics when designing human-AI interaction experiments. Future work could further explore mechanisms for validating variables such as AI suggestion explainability and task pressure, thereby optimizing human-AI collaboration in complex and dynamic scenarios.