Answers are grounded in this paper and cite their sources.
Artificial intelligence (AI) can improve human judgement, but human–AI systems rarely achieve complementarity: combined performance exceeding that of both humans and AI alone. Across three longitudinal studies (N = 671), we tested whether large language models (LLMs) could support complementarity by contributing to the construction and revision of everyday decision representations. Participants reported real, personally meaningful decisions involving two options and the decision outcome success one week later. As decisions were participant-generated rather than predefined, LLMs dynamically generated considerations and alternative evaluations tailored to each decision. We compared the predictive validity of AI, Human, and augmented evaluations. AI evaluations were standalone LLM predictions of decision outcome success. Human weighted-additive evaluations combined participant-generated attributes, importance weights, and option ratings. Augmented evaluations incorporated LLM-generated attributes and, in later studies, suggested weights and ratings, while participants selected attributes and retained control over evaluations. Augmented evaluations consistently predicted experienced outcomes and showed the strongest observed associations across all three studies. Evidence for complementarity was clearest in Study 2, where fully augmented evaluations outperformed both standalone approaches and explained additional variance in expected and experienced outcomes beyond the standalone evaluations considered together. Across studies, augmentation expanded the attribute set, with participants retaining human and AI contributions. These findings show that human–AI complementarity can emerge in predicting success of real decision outcomes and suggest that decision architecture matters. A distributed cognitive system that integrates AI contributions across all stages of decision making while humans remain responsible for selection and judgement of personally relevant considerations offers a promising architecture for evaluating diverse, personally meaningful real-life decisions.
No takes yet. Share an insight, caveat, or question.
Banks et al. (2026) studied this question.
Synapse has enriched 8 closely related papers on similar clinical questions. Consider them for comparative context: