Toward Resilient Human–AI Collaboration: A Lifecycle Taxonomy of Sociotechnical Risks and Cascading Failures

Human-AI collaboration is often analyzed through domain-specific lenses, such as clinical alert fatigue in healthcare or editorial accountability in journalism. This paper argues that these fragmented perspectives obscure a recurring set of underlying sociotechnical failure mechanisms. By adopting a lifecycle-oriented approach, the authors demonstrate that collaboration failures are not isolated technical glitches but are instead interconnected dynamics that evolve as humans and AI interact over time.

The authors map the collaboration process across four sequential stages:

1. Task Allocation: The initial division of labor, where issues like metaknowledge deficits (not knowing one's own limitations) and authority confusion can undermine the foundation of the partnership.
2. Interaction: The active phase where cognitive burden, explanation fatigue, and epistemic drift (the illusion of correctness) can lead to poor decision-making.
3. Feedback: The process of trust calibration, where inadequate feedback loops lead to either dangerous over-reliance or unwarranted algorithm aversion.
4. Adoption: The long-term integration phase, characterized by risks such as capability erosion (deskilling), technostress, and the widening of accountability gaps.

This taxonomy is significant because it highlights the "Explainable AI (XAI) paradox," where interventions intended to improve transparency—such as providing more information—can inadvertently increase cognitive load and reinforce automation bias. By identifying six recurring risk clusters—Trust Miscalibration, Cognitive Burden, Accountability Gap, Capability Erosion, Goal Misalignment, and AI Anxiety—the framework provides a roadmap for designers and policymakers to move beyond piecemeal interventions. Instead of fixing symptoms in isolation, this approach encourages governance and design strategies that address the cascading nature of sociotechnical risks throughout the entire lifecycle of an AI system.