The Interpretability of Artificial Intelligence and the Impact of Outcome Feedback on Trust: A Comparative Study
Dissertation Summary
summary
The researchers conducted two sets of experiments ("Predict the speed-dating outcomes and get up to $6 (takes less than 20 min)" and a similar Prolific experiment) in which participants interacted with the AI system in a task of predicting the outcome of a dating to explore the impact of model explainability and feedback on user trust in AI and prediction accuracy. The results show that although explainability (e.g., global and local interpretation) does not significantly improve trust, feedback can most consistently and significantly improve behavioral trust. However, increased trust does not necessarily lead to the same level of performance gains, i.e., there is a "trust-performance paradox". Exploratory analysis reveals the mechanisms behind this phenomenon.
Problem finding
The researchers found that although it is generally believed that the interpretability of the model can help improve the user's trust in the AI system, in the actual experiment, the global and local interpretability does not lead to a stable and significant trust improvement. Conversely, feedback (i.e., the output of the results) has a more significant effect on increasing user trust in the AI. However, this increased trust does not directly translate into an equivalent improvement in performance.
solution
To assess trust more accurately, the researchers used behavioral trust (WoA), a measure that takes into account the difference between the user's predictions and the AI's recommendations, and is independent of the model's accuracy. By comparing WoA under different conditions, researchers can analyze the relationship between trust and performance.
outcome
The results show that feedback has a more significant impact on improving users' trust in AI than explainability, but this enhanced trust does not lead to a corresponding performance improvement. Further exploration suggests that feedback induces users to over-trust (i.e., accept the AI's suggestions when it is wrong) or distrust (ignore the AI's suggestions when it is correct), which may negate the benefits of increased trust, leading to a "trust-performance paradox". The researchers call for future research to focus on how to design strategies to ensure that explanations foster appropriate trust to improve the efficiency of human-robot collaboration.
Draw inferences
Q1: How does feedback affect users' trust in AI?
A1: According to research, feedback (e.g. result output) is a key factor influencing user trust. It is the most significant and reliable way to increase user trust in AI behavior.
Q2: Does explainability necessarily enhance users' trust in AI?
A2: Although it is generally believed that the explanatory nature of the model helps to improve user trust, the experimental results show that this enhancement is not significant and not as effective as feedback. In specific cases, such as areas of low expertise, some form of interpretation may result in only a modest increase in appropriate trust.
Q3: How does result feedback and model interpretability affect user task performance?
A3: The study found that the feedback of the results can improve the accuracy of the user's predictions (reducing the absolute error), thereby improving the performance of working with AI. However, interpretability does not have as much impact on user task performance as it does on trust. This may mean that we should pay more attention to how to effectively use feedback mechanisms to improve the usefulness and effectiveness of AI-assisted decision-making.
Original address:https://dl.acm.org/doi/10.1145/3613904.3642780
The content is made up of:MiX CopilotBased on large language model generation, there may be a risk of errors.