Title
Human–AI Collaborative Decision-Making Under Uncertainty: A Conceptual Framework and Proposed Benchmark Specification
Authors
Abstract
Purpose — Human–AI collaboration under uncertainty remains under-theorized across policy, energy, finance, crisis response, and strategic planning. This conceptual article develops HACDU (Human–AI Collaborative Decision-Making Under Uncertainty), a framework that integrates Bayesian decision theory, bounded rationality, trust dynamics, explanation policies, and conflict resolution, and proposes HACDU-Bench, a benchmark specification for future cross-domain evaluation.
Design/methodology/approach — The study is conceptual and combines literature synthesis, formal modeling, and a design-science approach. Collaboration is modeled as a partially observable, utility-coupled stochastic process, from which a five-layer core decision pipeline, optional extensions, and five testable propositions are derived. No human-subject data were collected, and no confirmatory experiment was conducted.
Findings — As a conceptual study, the article reports theoretical rather than empirical findings. It offers (C1) a unified formalism centered on a core pipeline of belief fusion, trust dynamics, and override control; (C2) a modular algorithmic stack that makes its two strongest assumptions operational through a post-hoc calibration wrapper for AI uncertainty and a behavioral-telemetry estimator of human reliance; and (C3) HACDU-Bench as a proposed benchmark specification with deployment prerequisites. A characterization of positive collaborative value added (CVA) and five testable propositions complete the framework. A single energy-dispatch example shows how belief fusion, trust-dependent thresholds, value aggregation, and override logic interact in one decision cycle; it illustrates the mechanics of the framework and is not empirical validation.
Research limitations/implications — The propositions remain untested. Human-subject experiments, cross-domain replication, and field deployment are needed before any claim about performance can be made.
Practical implications — Once validated, HACDU could serve as a template for designing, auditing, and governing human–AI decision systems in policy analysis, energy operations, finance, and crisis management. HACDU-Bench offers a transparent structure for evaluating collaborative rather than purely predictive performance.
Originality/value — The article treats the collaborative human–AI system, rather than the standalone model, as the unit of analysis. It unifies decision theory, explainable AI, human factors, calibration, and benchmarking in one formal architecture and sets out a research agenda for hybrid intelligence.
Keywords
Human–AI collaboration; decision-making under uncertainty; Bayesian belief fusion; post-hoc calibration; explainable AI; collaborative value added
Classification-JEL
C44, D81, D91, G11, G41, M51
Pages
73-108
How to Cite
Chan, L. C. W. J., & Pan, S. H. (2026). Human–AI Collaborative Decision-Making Under Uncertainty: A Conceptual Framework and Proposed Benchmark Specification. Annals of Artificial Intelligence, 2026(1), 73-108.