"It’s Not the AI’s Fault Because It Relies Purely on Data": How Causal Attributions of AI Decisions Shape Trust in AI Systems

Research Background and Issues

Identified Problems and Challenges

• The authors point out that the issue of how causal attribution behind AI decision-making affects trust in AI has not been thoroughly studied. Traditional trust calibration methods, such as providing explanations and confidence scores, fail to adequately consider users' perceptions of causal attribution in AI decisions.
• While AI is often seen as a tool for reducing costs and improving efficiency, its potential for errors raises critical questions about when users should trust AI. Particularly, when AI decisions are perceived as autonomous or influenced by external factors, users may assign varying degrees of responsibility or trust to the AI.

Importance of the Issue

• Insufficient or excessive trust in AI decision-making can have negative impacts on practical applications, especially in high-risk domains (e.g., medical diagnosis) or low-risk domains (e.g., music recommendation), where the effects may be more pronounced.
• The issue of responsibility allocation for AI errors prompts important considerations about the appropriateness of trust—for instance, whether users are more inclined to trust an AI that errs due to poor external data quality rather than one that fails due to internal algorithmic flaws.

Research Motivation and Related Work

• Previous studies have primarily focused on enhancing AI transparency to improve trust calibration (e.g., through explainable AI). However, these studies have rarely explored how users understand causal attribution in AI decision-making.
• The researchers aim to systematically investigate how causal attribution (internal attribution or external attribution) influences trust in AI, as well as how decision stakes and outcome favorability moderate this relationship.

Solution

Methods and Design

• The study proposes an experimental design manipulating three independent variables:
  1. Causal Attribution: Signals indicating whether the causal attribution of AI decisions is internal (stemming from the AI itself) or external (dependent on external factors such as data quality).
  2. Decision Stakes: Differentiating between high-risk scenarios (e.g., medical diagnosis) and low-risk scenarios (e.g., music recommendation).
  3. Outcome Favorability: Distinguishing between decisions that are favorable or unfavorable to participants.
• Using scenario-based experiments, the study induces different perceptions of attribution and collects situational trust scores and open-ended responses from 192 participants.

Innovations

• Emphasizing causal attribution as a core factor influencing trust in AI—a dimension rarely directly explored in trust research.
• Investigating the interaction effects between causal attribution and contextual factors (decision stakes and outcome favorability), addressing gaps in the AI trust research field.
• Employing Kelley’s causal attribution framework (three informational variables: consensus, distinctiveness, consistency) to precisely operationalize causal attribution, ensuring high internal validity in the experiments.

Research Findings

Experimental Results

1. Main Effect of Causal Attribution:

   • Participants exhibited significantly higher trust when AI decisions were attributed to external factors compared to internal factors.
   • Internal attribution heightened perceptions of AI autonomy and responsibility, thereby reducing trust; external attribution framed AI as a tool reliant on external data, enhancing trust.

2. Interaction Effects Between Causal Attribution and Contextual Factors:

   • In low-risk scenarios, external attribution had a more pronounced effect on increasing trust compared to internal attribution.
   • In high-risk scenarios, while external attribution still boosted trust, participants generally exhibited lower trust in high-risk AI scenarios, indicating that risk perception plays a dominant role in these contexts.

3. Independent Effect of Outcome Favorability:

   • Favorable outcomes significantly increased trust, whereas unfavorable outcomes reduced trust.
   • However, outcome favorability did not significantly moderate the effect of causal attribution on trust.

Advantages Over Existing Solutions

Compared to existing trust calibration methods (e.g., confidence scores or explanations), this approach focuses on the critical role of causal attribution in user understanding, capturing cognitive mechanisms in trust formation more effectively and providing targeted optimization for human-AI interaction design.

Limitations and Future Work

• Limitations:
  • The study only examined two extreme cases of causal attribution (internal/external) without considering shared or dynamic causal scenarios.
  • The experiments involved hypothetical AI systems and did not account for variables introduced by real-world AI applications.
  • The interaction between AI transparency or explainability design and causal attribution perception was not explored.
• Future Directions:
  • Investigating dynamic changes in causal attribution within shared decision-making or collaborative scenarios.
  • Validating the trust model's applicability with user experiences in real-world AI systems.
  • Exploring trust calibration strategies for high-risk scenarios, such as combining causal attribution with enhanced transparency measures or human oversight.

Conclusion and Implications

This study is the first to explicitly identify the central influence of causal attribution on trust in AI and its contextual moderation characteristics. The research suggests:

1. Clearly signaling the causal mechanisms behind AI decisions in human-AI interactions, such as describing AI as a data-dependent tool reliant on external inputs.
2. Developing specific trust calibration strategies for high-risk scenarios, such as combining causal attribution with methods to enhance transparency.
3. Avoiding excessive portrayal of AI systems as autonomous and intelligent to mitigate unrealistic expectations and skepticism.

By more effectively communicating the causal nature of AI decision-making, this study provides theoretical support and practical guidance for fostering appropriate trust (rather than blind trust) in AI systems.