This site is a work in progress and has not been widely shared. Content may contain errors. Feedback is welcome.
This site is undergoing review. Some annotations were human-generated, some AI-generated — all are being verified.
Back to papers

Explanations Can Reduce Overreliance on AI Systems During Decision-Making

Vasconcelos, Jörke, Grunde-McLaughlin, Gerstenberg, Bernstein, Krishna

2023Proceedings of the ACM on Human-Computer Interaction219 citations
Experimental evidenceComputer Science / AICausalTheoretical model
Machine Learning (pre-LLM)Human-AI collaborationDecision-makingAugmentation vs. substitution
Abstract

Prior work has identified a resilient phenomenon that threatens the performance of human-AI decision-making teams: overreliance, when people agree with an AI, even when it is incorrect. Surprisingly, overreliance does not reduce when the AI produces explanations for its predictions, compared to only providing predictions. Some have argued that overreliance results from cognitive biases or uncalibrated trust, attributing overreliance to an inevitability of human cognition. By contrast, our paper argues that people strategically choose whether or not to engage with an AI explanation, demonstrating empirically that there are scenarios where AI explanations reduce overreliance. To achieve this, we formalize this strategic choice in a cost-benefit framework, where the costs and benefits of engaging with the task are weighed against the costs and benefits of relying on the AI. We manipulate the costs and benefits in a maze task, where participants collaborate with a simulated AI to find the exit of a maze. Through 5 studies (N = 731), we find that costs such as task difficulty (Study 1), explanation difficulty (Study 2, 3), and benefits such as monetary compensation (Study 4) affect overreliance. Finally, Study 5 adapts the Cognitive Effort Discounting paradigm to quantify the utility of different explanations, providing further support for our framework. Our results suggest that some of the null effects found in literature could be due in part to the explanation not sufficiently reducing the costs of verifying the AI's prediction.

Summary

Vasconcelos et al. conduct five lab experiments with 731 crowdworkers using a maze-solving task to test a cost-benefit framework predicting when AI explanations reduce overreliance on incorrect predictions in human-AI decision-making.

Main Finding

Explanations reduce overreliance on incorrect AI predictions when tasks are difficult and explanations are easy to understand; in hard tasks, highlight explanations reduced overreliance by 1.74 percentage points compared to prediction-only; written explanations (harder to parse) showed no reduction; higher monetary bonuses reduced overreliance by 0.93 percentage points

Primary Datasets

Experiment-generated data from five lab studies on Prolific platform

Secondary Datasets

None

Key Methods
Five lab experiments manipulating task difficulty (maze size: 10x10, 25x25, 50x50), explanation modality (prediction-only, highlight, written, incomplete, salient), and monetary bonuses; Bayesian linear mixed effects models; Cognitive Effort Discounting (COG-ED) paradigm to measure subjective utility
Sample Period
2021-2022
Geographic Coverage
United States
Sample Size
731 participants across 5 studies (Study 1: N=340, Study 2: N=340, Study 3: N=286, Study 4: N=114, Study 5: N=76)
Level of Analysis
Individual
Occupation Classification
None
Industry Classification
None
Notes
Proceedings of the ACM on Human-Computer Interaction, vol. 7, pp. 1-38 [Claude classification]: Published in Proceedings of the ACM on Human-Computer Interaction, vol. 7, pp. 1-38. Study uses simulated AI (not real ML model) with controlled accuracy of 80%. Explanations are programmatically generated to be accurate when AI is correct and to show errors when AI is incorrect. Pre-registered studies (osf.io/vp749, osf.io/g6tjh, osf.io/4dbqp, osf.io/hgz2x, osf.io/cskvb). [Claude classification]: Published in Proceedings of the ACM on Human-Computer Interaction, vol. 7, pp. 1-38. Study uses simulated AI (not real ML model) with controlled accuracy of 80%. Explanations are programmatically generated to be accurate when AI is correct and to show errors when AI is incorrect. Pre-registered studies (osf.io/vp749, osf.io/g6tjh, osf.io/4dbqp, osf.io/hgz2x, osf.io/cskvb). [Claude classification]: Published in Proceedings of the ACM on Human-Computer Interaction, vol. 7, pp. 1-38. Study uses simulated AI (not real ML model) with controlled accuracy of 80%. Explanations are programmatically generated to be accurate when AI is correct and to show errors when AI is incorrect. Pre-registered studies (osf.io/vp749, osf.io/g6tjh, osf.io/4dbqp, osf.io/hgz2x, osf.io/cskvb). [Claude classification]: Published in Proceedings of the ACM on Human-Computer Interaction, vol. 7, pp. 1-38. Study uses simulated AI (not real ML model) with controlled accuracy of 80%. Explanations are programmatically generated to be accurate when AI is correct and to show errors when AI is incorrect. Pre-registered studies (osf.io/vp749, osf.io/g6tjh, osf.io/4dbqp, osf.io/hgz2x, osf.io/cskvb). [Claude classification]: Published in Proceedings of the ACM on Human-Computer Interaction, vol. 7, pp. 1-38. Study uses simulated AI (not real ML model) with controlled accuracy of 80%. Explanations are programmatically generated to be accurate when AI is correct and to show errors when AI is incorrect. Pre-registered studies (osf.io/vp749, osf.io/g6tjh, osf.io/4dbqp, osf.io/hgz2x, osf.io/cskvb). [Claude classification]: Published in Proceedings of the ACM on Human-Computer Interaction, vol. 7, pp. 1-38. Study uses simulated AI (not real ML model) with controlled accuracy of 80%. Explanations are programmatically generated to be accurate when AI is correct and to show errors when AI is incorrect. Pre-registered studies (osf.io/vp749, osf.io/g6tjh, osf.io/4dbqp, osf.io/hgz2x, osf.io/cskvb). [Claude classification]: Published in Proceedings of the ACM on Human-Computer Interaction, vol. 7, pp. 1-38. Study uses simulated AI (not real ML model) with controlled accuracy of 80%. Explanations are programmatically generated to be accurate when AI is correct and to show errors when AI is incorrect. Pre-registered studies (osf.io/vp749, osf.io/g6tjh, osf.io/4dbqp, osf.io/hgz2x, osf.io/cskvb). [Claude classification]: Published in Proceedings of the ACM on Human-Computer Interaction, vol. 7, pp. 1-38. Study uses simulated AI (not real ML model) with controlled accuracy of 80%. Explanations are programmatically generated to be accurate when AI is correct and to show errors when AI is incorrect. Pre-registered studies (osf.io/vp749, osf.io/g6tjh, osf.io/4dbqp, osf.io/hgz2x, osf.io/cskvb). [Claude classification]: Published in Proceedings of the ACM on Human-Computer Interaction, vol. 7, pp. 1-38. Study uses simulated AI (not real ML model) with controlled accuracy of 80%. Explanations are programmatically generated to be accurate when AI is correct and to show errors when AI is incorrect. Pre-registered studies (osf.io/vp749, osf.io/g6tjh, osf.io/4dbqp, osf.io/hgz2x, osf.io/cskvb). [Claude classification]: Published in Proceedings of the ACM on Human-Computer Interaction, vol. 7, pp. 1-38. Study uses simulated AI (not real ML model) with controlled accuracy of 80%. Explanations are programmatically generated to be accurate when AI is correct and to show errors when AI is incorrect. Pre-registered studies (osf.io/vp749, osf.io/g6tjh, osf.io/4dbqp, osf.io/hgz2x, osf.io/cskvb). [Claude classification]: Published in Proceedings of the ACM on Human-Computer Interaction, vol. 7, pp. 1-38. Study uses simulated AI (not real ML model) with controlled accuracy of 80%. Explanations are programmatically generated to be accurate when AI is correct and to show errors when AI is incorrect. Pre-registered studies (osf.io/vp749, osf.io/g6tjh, osf.io/4dbqp, osf.io/hgz2x, osf.io/cskvb).