by Anna Kawakami
Kawakami, A., Taylor, J., Fox, S., Zhu, H., & Holstein, K. (2026). AI failure loops in devalued work: The confluence of overconfidence in AI and underconfidence in worker expertise. Big Data & Society, 13(1), 20539517261424164.
From headlines to prediction markets, AI’s impact on work is increasingly framed as a question of which jobs will disappear. Yet, in many workplaces, the more immediate issue is not displacement but failure. In historically devalued fields such as K-12 teaching, social work, and home health care, AI systems are often built on reductionist understandings of human work, leading to deployments that fall short in patterned ways. We describe these recurring breakdowns as AI Failure Loops: dynamics in which the devaluation of worker expertise shapes flawed AI systems—and those systems, in turn, further erode recognition of that expertise.
We conduct a focused review of academic and grey literature on AI deployments across three domains of devalued, feminized labor 1 in the United States (social work, home health, K-12 teaching), grounded in our team’s direct experiences studying and designing AI alongside workers in these fields. Through three case studies—AI-based risk assessments in child welfare, AI for home healthcare, and AI tutoring systems in K-12 teaching—we illustrate how AI Failure Loops arise in practice. We identify a set of six interconnected failure modes in the design, development, and deployment of AI systems that underlie AI Failure Loops: Expertise Misunderstanding, Managerial Over Worker Needs, Design Exclusion, Inappropriate Evaluation, Forced Use, and Unwarranted Blame (Figure 1).
Figure 1: A visual overview of AI Failure Loops. The six failure modes (in circles) that contribute to the dynamics of the AI Failure Loops exist within a web to illustrate the inter-connected relationship amongst the failure modes.
These failures are not isolated problems, but arise from a broader tension: AI Failure Loops emerge from the confluence of overconfidence in the capabilities of AI systems and underconfidence in the skills and expertise of workers. Today, with recent waves of both AI hype and actual increases in AI capabilities, developers are targeting a rapidly expanding range of socially complex tasks across feminized and other devalued occupations. Drawing lessons from past AI deployments in these contexts, we argue that this may further accelerate gross underestimations of worker expertise and overestimations of AI capabilities.
Importantly, nothing about the future of work is inevitable. By genuinely centering and uplifting workers in the design, evaluation, and governance of AI systems, we can begin to reverse these loops and develop AI that actually supports and dignifies, rather than diminishes, worker capabilities. Our paper ends with implications for shifting towards this future of pro-worker AI—a future where worker expertise is celebrated and uplifted through AI practice, and by society more broadly.
1 Feminized labor is a particularly extreme example of devalued labor; historically, it has been mislabeled “women’s work,’’ and today, these jobs continue to be predominantly performed by women and people of color. Workers in these fields frequently report feeling undervalued and overworked, with many roles experiencing high retention rates.
