Working Paper
Organizations increasingly adopt AI tools in decision-making, yet humans often exhibit AI aversion and disregard its inputs. While granting managers full autonomy in weighting AI recommendations can address this aversion, it risks limiting potential gains if not used effectively, leading to lower performance. Through an online experiment where participants are tasked with making collaborative predictions leveraging AI advice, the authors examine this autonomy-performance trade-off.
Their results reveal that participants with full autonomy exhibit systematic initial weight inertia characterized by anchoring to initial AI weights and a ceiling effect where users fail to exceed initial weights put on AI advice. Participants also disregard performance feedback, failing to adjust weights even when consistently underperforming compared to AI. To mitigate this inertia while preserving autonomy, the authors develop a bounded autonomy approach, constraining weight adjustments based on collaborative performance relative to AI’s predictions alone. When AI performs well, participants cannot fully disregard its recommendations, especially when their own predictions are inaccurate. This approach reduces human-AI collaborative prediction errors by 12.47%, reducing initial weight inertia and improving feedback responsiveness.
Finally, the authors find that different framings (weighting AI vs. weighting participants’ own predictions) yield similar performance improvements but via distinct mechanisms: weighting AI predictions increases cognitive effort, while weighting original predictions eliminates anchoring bias. This novel approach balances user autonomy with improved performance, offering a practical solution for enhancing human-AI collaboration.
