Do people engage cognitively with AI? Impact of AI assistance on incidental learning

Title of the Paper

Do People Engage Cognitively with AI? Impact of AI Assistance on Incidental Learning

Paper Information

• Authors: Krzysztof Z. Gajos and Lena Mamykina
• Publication Date and Venue: 2022, 27th International Conference on Intelligent User Interfaces (IUI '22)
• Thematic Areas: Artificial Intelligence and User Interaction, Cognitive Psychology, Educational Learning
• Keywords: Decision Support Systems, Incidental Learning, Cognitive Engagement, Explainable AI, Human-Centered AI

Research Background and Problem

• Identified Problems or Challenges:

  • Modern AI provides decision-making advice to humans, but users often process AI-provided information superficially.
  • When collaborating with AI, people tend to blindly rely on AI, reducing deep cognitive processing of recommendations.
  • It remains unclear how AI-generated recommendations and explanations influence users' deep engagement and whether they can promote knowledge acquisition.

• Significance of the Research:

  • In high-stakes domains like healthcare, it is critical for users to deeply evaluate and integrate AI recommendations.
  • Optimizing AI-assisted decision-making systems through more effective interaction designs can enhance users' cognitive engagement and learning outcomes.

• Research Motivation and Related Work:

  • Previous studies suggest that merely providing AI recommendations and explanations may not be sufficient to promote deep user thinking.
  • "Incidental learning" (learning as a byproduct of other activities) is common in professional settings, but its effects in AI-assisted decision-making require further exploration.

Solution

• Research Objectives:

  • Investigate the impact of different forms of human-AI interaction on users' incidental learning.
  • Design and test three interaction formats:
    1. AI recommendations with explanations
    2. User decision-making followed by AI recommendations/explanations (“updated design”)
    3. AI explanations only (no specific recommendations provided)

• Innovative Contributions:

  • Systematic comparison of different interaction formats in terms of immediate benefits (improved decision quality) and learning outcomes (knowledge acquisition).
  • Emphasis on encouraging users to actively engage with AI interactions to foster deeper learning.

• Implementation Steps and Technical Methods:

  • Experimental Setup:
    • Design nutrition-related decision tasks (e.g., choosing which of two meals contains more of a specific nutrient).
    • Test participants' prior knowledge (pre-test), immediate benefits of interaction (intervention phase), and actual learning (post-test).
  • Experimental Conditions:
    • Compare three interaction designs with two baseline conditions (feedback only and feedback with explanations).
    • Measure normalized changes in immediate benefits and learning outcomes.
  • Data Collection and Analysis:
    • Collect experimental data via LabintheWild and Amazon MTurk platforms.
    • Analyze experimental data using non-parametric statistics (e.g., Wilcoxon signed-rank test).
    • Validate results for reproducibility.

Research Findings

• Specific Results:

  • Immediate Benefits:
    • Providing AI recommendations and explanations significantly improved participants' task performance but did not significantly promote learning.
    • The "explanations only" design supported immediate decision-making while significantly enhancing learning outcomes.
  • Learning Outcomes:
    • Participants learned new knowledge through the "explanations only" design, whereas other conditions (e.g., updated design) showed no significant differences in promoting learning.
  • The results indicate that requiring users to derive decisions from AI explanations is crucial for fostering deeper cognitive processing.

• Advantages Over Existing Approaches:

  • Interaction formats that provide explanations without explicit recommendations stimulate deeper user thinking and critical evaluation.
  • Compared to simple recommendation-plus-explanation approaches, this method is more effective in facilitating incidental learning.

• Experimental Evaluation and Reliability:

  • Replicated experiments confirmed the initial findings, demonstrating the generalizability of the design that guides users to make autonomous decisions based on AI explanations.
  • Effects were further analyzed for specific user groups (e.g., those with higher "need for cognition" tendencies).

• Limitations and Future Directions:

  • Limitations:
    • AI recommendations were always correct, so the sensitivity of explanation-guided decision-making to incorrect AI information was not evaluated.
    • The experimental tasks were low-risk, and the results may not directly generalize to high-stakes professional domains.
  • Future Directions:
    • Explore the potential of AI explanations in non-comparative scenarios (e.g., multi-option decision-making).
    • Investigate the impact of visual design on the depth of user processing of AI-generated information.
    • Further explore ways to reduce inequities caused by AI interventions across diverse populations.

This study highlights potential directions for optimizing current AI decision support designs by improving human-AI interaction models to foster deeper cognitive engagement and learning.