When do human–AI combinations achieve complementarity?

Determine the specific conditions and collaboration designs under which human–AI combination strategies achieve complementarity, defined as producing higher decision accuracy than both human-only and AI-only baselines, across different tasks and domains.

Background

The paper contrasts the dominant AI-as-advisor workflow with the hybrid confirmation tree (HCT), emphasizing how independence between human and AI judgments can improve accuracy. Despite these results, the broader literature lacks a clear understanding of when and which human–AI collaboration structures outperform both humans and AI individually.

Identifying the conditions for complementarity involves specifying task characteristics, relative accuracies of humans and AI, interaction structures (e.g., independent vs. advisory), and decision aggregation rules that yield combined performance gains beyond either component alone.

References

It is still largely unknown when and which human--AI combinations manage to achieve complementarity---that is, when they outperform both humans and AI \citep{vaccaro2024combinations}.

— Beyond AI advice -- independent aggregation boosts human-AI accuracy  (2603.29866 - Berger et al., 31 Mar 2026) in Discussion

Direct measurement of complementarity, reliance, delegation, and team performance therefore represents a substantial open area.

— What Counts as Strategic Reasoning? A Systematic Mapping of Chess Research on Humans, Engines, and Language Models  (2609.18286 - Ciancarini et al., 16 Sep 2026) in Section 4.4, “The collaboration gap”

Whether advanced chess achieved the functional interdependence its design intended, or settled into computer-assisted play with the human still deciding alone, is an empirical question rather than a definitional one.

— What Counts as Strategic Reasoning? A Systematic Mapping of Chess Research on Humans, Engines, and Language Models  (2609.18286 - Ciancarini et al., 16 Sep 2026) in Section 5.3, “Hybridization Operates at Different Levels”

Whether this helps students learn to collaborate with AI remains a question for the studies described in Section~\ref{sec:future_work}.

— DataCanvas-EDU: An Agentic Framework for Instructor-Guided Synthetic Data Generation in Business Analytics Education  (2609.19617 - An et al., 17 Sep 2026) in Section 2.3, “AI-assisted analysis and training data contamination” (cross-referenced to Section 6, “Future Work”)

This overall comparison does not establish synergy or equivalence.

— Confident, Not Wiser: The Dunning-Kruger Effect in Human-AI Interaction  (2609.31095 - Fernandes et al., 25 Sep 2026) in Section 5, subsection “Human-AI Joint Performance”; Discussion, subsection “Assessing a partnership, not only oneself”

While technical research on MASS has progressed rapidly , human-centered integration remains an open challenge .

— Understanding Operator Attitudes Toward AI-Supported Decision Making in Maritime Operations  (2609.11805 - Jirak et al., 10 Sep 2026) in Section 1, Introduction