A Dynamic Human–AI Decision-Making Model for Cognitive Adaptation, Algorithmic Reliance, and Decision Optimization

Trust, confidence, and reliance on algorithmic recommendations are integral to human–AI decision systems, which are increasingly reliant on how well people calibrate these constructs in relation to algorithmic recommendations. This study analyzed a human-in-the-loop decision dataset of 10,227 reviewed cases to explore the nature of cognitive adaptation, algorithmic dependence, override behaviors and decision optimization. The analysis was limited to the accuracy of human participants, machine-learning accuracy, agreement between human and machine learning, level of confidence, decision latency, expertise level, ML-confidence quartiles, override outcomes and chronological stability. The overall accuracy of the AI system was 97.27% and 90.21% for the human reviewers, and the human–AI agreement was 88.01%. The expert reviewers were highly associated with better performance, and the results showed that the expert reviewers' accuracy, decision time, confidence and agreement with AI were better than the intermediate and novice reviewers. 20.55% of 1226 overrides by humans were to correct AI errors, while 79.45% were made by humans to replace correct AI predictions with incorrect human ones. There was no monotonic increase in agreement as a function of ML confidence, with a decrease to 81.70% in the highest confidence quartile. The overall pattern of decision indicators was also consistent over time. The results indicate that interventions that depend on human expertise and are calibrated to the human agent's level of confidence are appropriate for optimized human–AI decision systems, rather than human override or algorithmic reliance in all situations.