Contrastive Explanations That Anticipate Human Misconceptions Can Improve Human Decision-Making Skills

Research Background and Issues

• Identified Problems or Challenges: The study reveals that although AI systems provide decision support and explanations, users do not significantly learn from these explanations or may rely on AI excessively, leading to skill degradation. When AI recommendations are incorrect, people still tend to over-rely on them. This issue is particularly severe as multiple studies show that many individuals naturally prefer "contrastive explanations" rather than unidirectional ones, which are typically based solely on AI decision logic and do not account for users' knowledge or thought processes.
• Significance: The widespread application of AI may lead to critical skill degradation, especially in high-risk decision-making domains such as healthcare and finance, where users need strong independent thinking abilities to identify and assess AI errors. This not only impacts long-term autonomy but may also negatively affect the development of professional skills.
• Research Motivation and Related Work: Previous research on contrastive explanations often defines the "contrastive item" primarily as alternative categories within the model, without considering users' actual mental models. Furthermore, related studies have demonstrated that users under unidirectional explanation conditions are more likely to exhibit over-reliance, reducing their ability to learn from tasks. Therefore, research needs to explore a new explanatory framework to better facilitate human learning.

Solution

• Proposed Solution or Method: The authors propose a framework for generating human-centered contrastive explanations, which highlight the differences between AI choices and predictions and human choices (i.e., "Why choose A instead of B"). Here, "B" is derived from the user's mental model as a potential choice, rather than merely an alternative category selected by the model.
• Innovative Aspects: This framework emphasizes generating explanations by predicting the contrastive items users might choose (foils), avoiding direct reliance on user input or simplifying the explanation to unidirectional alternative categories. The framework contrasts AI recommendations with common human choices, highlighting the knowledge differences between the two.
• Implementation Steps and Key Technologies:
  1. Module Design:
     • AI Task Model: Provides AI decision recommendations (referred to as facts).
     • Human Model: Predicts possible user choices based on user data (referred to as contrastive items).
     • Contrastive Module: Analyzes the differences between AI recommendations and human choices, marking which dimensions are superior or inferior.
     • Presentation Module: Uses large language models (LLMs) to generate clear and comprehensible natural language explanations, supplemented with necessary commonsense knowledge.
  2. Experimental Design: Constructs explanations for a hypothetical sports advice task and designs five experimental conditions (no AI, unidirectional explanation, predicted contrastive items, random contrastive items, and input-based contrastive items).

Research Outcomes

• Specific Outcomes:
  • Using contrastive explanations (especially those generated from predicted contrastive items) significantly enhances participants' learning ability, outperforming conditions with no AI support or unidirectional explanations.
  • Providing contrastive explanations based on user mental models does not reduce decision accuracy.
  • For contrastive explanations (predicted vs. randomly generated), high-quality predicted contrastive items further improve user learning.
• Advantages Over Existing Solutions:
  • Compared to unidirectional explanations, contrastive explanations better align with users' mental models, making it easier for them to understand knowledge gaps.
  • Offers personalized designs tailored to users' knowledge states, helping optimize their learning experience.
• Experimental or Evaluation Results:
  • Learning Outcomes: Learning under the contrastive explanation (predicted contrastive items) condition was significantly better than under the unidirectional explanation condition (p < 0.05).
  • Accuracy: Decision accuracy under contrastive and unidirectional explanation conditions was similar (no significant difference).
  • User Subjective Experience: Users found predicted contrastive item explanations to provide higher autonomy and trust in the system, whereas the subsequent "contrastive criticism" mode might reduce satisfaction with the task.
• Limitations and Future Directions:
  • Limitations: The experimental task was relatively simple, with short-term learning effects observed; long-term impacts were not studied. Additionally, simulated AI error recommendations may not fully represent real-world environments.
  • Future Directions: Further exploration of personalized contrastive item design by refining user mental models to optimize explanation quality; extending the framework to more complex multi-category decision scenarios; studying the effectiveness of explanation design in promoting long-term knowledge retention.

Conclusion

This study demonstrates that contrastive explanations generated based on user mental models can effectively enhance human independent decision-making skills while maintaining task accuracy. The framework provides a solid foundation for further optimization of AI-human interaction, particularly in designing systems that improve user learning capabilities in collaborative tasks.