The trust crisis in artificial intelligence: AI hallucinations and human-AI collaboration

The prevalence of AI hallucinations in generative artificial intelligence poses a serious challenge to human-AI collaboration, which relies heavily on trustworthy interactions. This phenomenon can lead to flawed decision-making and a crisis of trust. While existing research has highlighted the associated risks, there remains a lack of systematic empirical investigation and theoretical explanation regarding how these phenomena influence human-AI collaborative behavior via psychological mechanisms, as well as the nuanced role of task-technology fit (TTF) in this process. To address this gap, this study develops an integrated theoretical model incorporating human-machine trust and TTF, aiming to uncover the underlying pathways and boundary conditions through which AI hallucinations affect collaboration. Based on a survey of 310 users of AI-augmented work tools and analyzed using partial least squares structural equation modeling (PLS-SEM) and artificial neural network (ANN) techniques, the study revealed the following findings: AI hallucinations exert a significant negative impact on human-AI collaboration. Specifically, they impair collaborative outcomes indirectly by eroding users' cognitive and emotional trust in AI systems. Moreover, task-technology fit plays a bidirectional moderating role in this relationship. This study provides systematic empirical evidence quantifying the direct effect of AI hallucinations on human-AI collaboration. Furthermore, it advances theoretical understanding by identifying human-machine trust as a core mediating mechanism—thereby challenging the simplistic "hallucination→behavior" direct-path assumption. It also refines the conventional view of task-technology fit as a uniformly positive construct by uncovering its multifaceted roles in technology-related risk scenarios. These findings provide empirical grounding and practical insights for designing more robust and trustworthy human-AI collaborative systems. They also offer theoretical guidance for formulating effective hallucination risk management strategies and trust-calibration mechanisms in high-stakes application environments.