The Limits of Explanation in Human–AI Teaming: Evidence from Sports Forecasting

Author(s):

Ayobami Oyewole, Steven Thurman, and Ramesh Srinivasan

Abstract:

This study examines whether the relevance of AI explanations improves learning and performance in hybrid human–AI binary-event forecasting. Ultimate Fighting Championship outcomes were forecast by 45 participants across 50 trials while interacting with 1 of 3 conversational agents matched in accuracy (68%): GoodAI (highest-SHapley Additive exPlanations [SHAP] feature), BadAI (lowestSHAP feature), or NoAI (no rationale). Performance was incentivized wagering; learning was quantified as pre-/post-changes in subjective feature weights, compared with an optimal statistical model and with each condition’s rationale distribution. Contrary to expectations, AI condition had no reliable effect on final earnings (ANOVA F (2, 38) = 0.79, p = 0.46) or on alignment to optimal weights (F (2, 38) = 2.12, p = 0.13); rationale distributions showed no selective influence on behavior. There was a small trend favoring NoAI in terms of mental model alignment with the optimal feature weights. In feedback-rich tasks with compact feature sets, static single-feature rationales may add little informational value. Methodologically, we contribute a relevance-controlled evaluation paradigm for human-validated explainable AI.