인간-AI 조합에서 생성형 AI vs 전통적인 AI 간의 시너지효과 분석

This study investigates when and how different types of AI systems generate synergy in human–AI teams. Prior work shows a puzzling pattern: on average, human–AI combinations outperform humans alone but rarely exceed the better of the human or the AI partner. Strong synergy—teams doing better than both humans and AI alone—appears only in a minority of tasks, while mere augmentation of human performance is common. Against this backdrop, we ask whether AI system type (generative vs classical) moderates synergy and under what conditions generative AI (large language models, LLMs) is beneficial or harmful in human–AI collaboration. We re-analyze the meta-analytic dataset of human–AI team experiments compiled by Vaccaro et al. (2024), which aggregates 370 effect sizes from 106 studies comparing human solo, AI solo, and human–AI team performance on the same tasks. Excluding Woz-type systems, we use 305 effect sizes (8 generative AI, 297 classical AI). For each experimental condition, we compute standardized effect sizes (Hedges’ g) for four synergy metrics: strong synergy (g(HAI,max): team vs best solo), human augmentation (g(HAI,H): team vs human), AI augmentation (g(HAI,AI): team vs AI), and negative synergy (g(HAI,min): team vs worst solo). We then classify tasks by the relative advantage of AI over humans (RA>0: AI>Human; RA<0: Human>AI) and test five research questions comparing generative vs classical AI using Welch’s unequal-variance t-tests with Holm–Bonferroni correction. The results reveal a clear asymmetry. When AI outperforms humans (AI>Human), human–LLM teams achieve significantly higher strong synergy than human–classical AI teams: on average, they surpass the best solo performer, whereas classical AI teams fall below AI solo performance, indicating a loss from adding a human teammate. Generative AI teams also show positive AI-augmentation effects (team＞AI) in these AI-dominant tasks, while classical AI teams do not. Aggregated across all tasks, human–LLM teams display a positive mean strong synergy, whereas human–classical AI teams show a negative mean, suggesting that generative AI can raise the upper bound of human–AI performance. However, when humans outperform AI (Human>AI), the pattern reverses. Classical AI teams exhibit positive human-augmentation effects (team＞human), but generative AI teams, on average, reduce performance below the human baseline. Finally, in terms of negative synergy risk (team performing worse than the worst solo), classical AI teams more reliably avoid catastrophic failures than generative AI teams, whose performance is more volatile and occasionally falls below the weakest individual. These findings make three contributions. First, they provide empirical evidence that AI system type is a critical moderator of human–AI synergy: generative AI enables strong and AI-augmenting synergy primarily when AI is already better than humans, but may harm performance in expert-dominant tasks and does not systematically reduce worst-case outcomes. Second, they extend existing models of human–AI collaboration by highlighting a two-dimensional design space defined by relative human–AI advantage and AI type (generative vs classical), rather than treating “AI” as a single category. Third, they offer practical guidance for organizations: generative AI is best deployed for open-ended, creative, or high-complexity tasks where AI outperforms humans and can provide rich drafts or explanations, while classical AI is more suitable for supporting experts in narrow, safety-critical domains and for serving as a robust safeguard against catastrophic failure.