Position: Adopting AI in Practice Does Not Guarantee the Productivity Boost

Won Ik Cho, Seong-hun Kim, Geunhye Kim · 2026

Read Position: Adopting AI in Practice Does Not Guarantee the Productivity Boost by Won Ik Cho, Seong-hun Kim, Geunhye Kim (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2605-24688.txt.

Introduction

Won Ik Cho and others argue that adopting AI at work does not by itself guarantee higher productivity. Some sectors rushed to encourage AI use, while others restricted it. Most practitioners assume better technical capability means a productivity boost. The authors counter that human and organizational factors can weaken or even cancel the gain. They name five moderators: staff mix, baseline skill, learning curves, incentives for fair use, and flexibility of goals. They revise the economic frame and urge action on the human side.

Catalog record

CREATORWon Ik Cho, Seong-hun Kim, Geunhye Kim
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE48448
LICENSEhttp://creativecommons.org/licenses/by-sa/4.0/
RIGHTSCC BY SA
SOURCEarXiv

Source: https://arxiv.org/abs/2605.24688

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Position: Adopting AI in Practice Does Not Guarantee the Productivity Boost Won Ik Cho, Seong-hun Kim, Geunhye Kim Abstract This position paper argues that adopting AI in organizational practice does not guarantee productivity gains, because human and environmental factors critically moderate the relationship between AI deployment and realized productivity improvements . Following the advent of high-performance generative models, AI use has been rapidly encouraged in some sectors while being restricted in others. Most practitioners assume that AI brings productivity boosts owing to enhanced technical capabilities, but regardless of apparent performance advances in AI technology, human and environmental factors of the organization may substantially attenuate—or even negate—the effective productivity benefits. We identify five key moderating factors: human resource composition, baseline capability of individuals, learning curve of practitioners, incentives for fair use, and flexibility of objectives. Drawing on the partial equilibrium model of Gries and Naudé (2022), we argue that existing economic frameworks may inadvertently overlook these factors. We revise the existing framework to redefine effective organizational determinants and shed light on practical implications including industry and education, responding to alternative views and calling for action of stakeholders. 1 Introduction Artificial intelligence transformation represents one of the most significant technological shifts across modern society, affecting laypeople, practitioners, and domain experts alike ( 36 ) . The applications of AI have expanded dramatically in recent years, encompassing search and information retrieval ( 37 ) , recommendation systems ( 50 ) , personalized learning ( 34 ) , creative content generation ( 41 ) , and task automation ( 44 ) across virtually every sector. Decision-makers in organizations increasingly assume that adopting AI will boost productivity, and this assumption appears theoretically sound if all users possess adequate understanding of what AI systems can and cannot accomplish ( 7 ) . Recent research in economics has begun incorporating AI as a productivity factor in formal models of aggregate output. 26 provides empirical evidence on automation’s effects on labor productivity across industrialized economies, while 3 extends the task-based framework to compute AI’s contribution to total factor productivity. These approaches make intuitive sense under the assumption that AI capabilities translate directly into organizational performance gains. However, not all organizations achieve productivity improvements through AI introduction. A substantial body of evidence suggests that factors beyond technological capability determine whether AI adoption yields meaningful productivity gains. The “ productivity paradox ”—wherein AI capabilities advance rapidly while aggregate productivity growth remains sluggish—has been documented across multiple contexts ( 12 ) . Field experiments reveal substantial heterogeneity in AI’s effects: 11 found productivity gains of up to 36% for the lowest-skilled workers, with minimal impact on the most experienced staff, while 19 demonstrated that for tasks outside AI’s capability frontier, workers using AI performed 19 percentage points less likely to be accurate than those without AI access. Cross-country firm-level evidence shows that AI adoption is highly concentrated among large, already-productive firms, with complementary assets—including skilled workers, organizational processes, and data infrastructure—playing a critical moderating role ( 14 ) . 3 estimates that AI’s contribution to total factor productivity may be substantially lower than commonly assumed, with effects highly dependent on the share of tasks that are easy- vs. hard-to-learn, and that whether feasibly automatable within the relevant horizon. The gap between AI’s technical potential and realized organizational benefits echoes Solow’s famous observation that “ you can see the computer age everywhere but in the productivity statistics ” ( 42 ) . We take the position that adopting AI in practice does not guarantee productivity boosts, because human and environmental factors—including organizational structure, individual capabilities, learning dynamics, usage incentives, and goal flexibility—critically affect the relationship between AI deployment and realized productivity gains. Traditional perspectives that treat productivity improvements as automatic consequences of AI capability adoption are thus insufficient and potentially lacking detail for practitioners. We identify five human and environmental factors that moderate AI productivity effects: (1) human resource composition of the organization, (2) baseline capability of individuals who interact with AI, (3) learning curve that practitioners experience, (4) incentives for fair and appropriate AI use, and (5) flexibility of objectives and key results. We argue that the economic model of 27 , while valuable, treats these factors as exogenous parameters rather than endogenous organizational variables that practitioners must actively manage. We will also see this productivity gap manifests differently across domains with high AI utilization, briefly investigating cases of industry and education. 2 Background 2.1 Recent Advances in High-Performance AI The past several years have witnessed remarkable advances in AI capabilities, driven by a series of architectural and methodological breakthroughs. The scaling of autoregressive language models culminated in GPT-3 ( 10 ) , which exhibited remarkable few-shot learning capabilities without task-specific fine-tuning. For coding, specialized models such as Codex ( 17 ) demonstrated the ability to generate functional code from natural language descriptions, powering tools like GitHub Copilot. The introduction of instruction-tuning and reinforcement learning from human feedback ( 38 ) transformed language models into conversational assistants capable of following complex user instructions, exemplified by systems like ChatGPT. Most recently, agentic AI frameworks that integrate reasoning and action-taking ( 48 ) have extended language models to autonomous task completion through interaction with external tools and environments. These technical advances have collectively fueled widespread interest in AI adoption across sectors. 2.2 Human Perception and Productivity Research Research at the intersection of human-computer interaction and computational social science has examined how humans perceive and interact with AI systems. 24 provides a comprehensive review of empirical research on human trust in AI, identifying factors including system transparency, reliability, and anthropomorphism that shape user attitudes and behaviors. 46 surveys methodological approaches for studying trust in AI-assisted decision making, highlighting the complexity of human responses to AI assistance. Empirical studies of AI’s effects on work productivity have begun to emerge. 11 conducted a large-scale field experiment with customer support agents at a software company, finding that generative AI access increased productivity by 15% on average, with substantially larger gains ( \sim 36%) for novice workers compared to experienced staff. 19 introduced the concept of a “ jagged technological frontier ,” demonstrating through a randomized experiment with management consultants that AI improves performance for tasks within its capabilities while degrading performance for tasks outside that frontier. These findings suggest that productivity effects are highly contingent on task characteristics and worker attributes. In education sector, 49 systematically reviewed AI applications in higher education, finding that research has predominantly focused on technical implementations rather than pedagogical considerations. More recent work examining large language model adoption ( 47 ) identified practical and ethical challenges including academic integrity concerns, hallucination risks, and equity implications. 20 conducted a meta-analysis of experimental studies on ChatGPT in education, finding positive effects on academic performance but no significant improvement in student self-efficacy. 2.3 Economic Models of AI and Productivity The theoretical and economic modeling of AI’s productivity effects has also drawn attention. The task-approach to labor markets, developed by 5 and extended by 1 , provides a foundation for understanding how technologies affect the allocation of tasks between human workers and machines. 2 incorporated this approach into an endogenous growth model, showing that automation has both displacement effects (substituting capital for labor in existing tasks)

Read the complete text