The Visibility Trap: Attention Structures and the Hidden Costs of Generative AI in Work and Governance

Technological advancement can open new possibilities for organizations, but whether and how these translate into organizational effects depends on how they are made visible, evaluated, and prioritized. Organizations do not engage in this process neutrally. Instead, they rely on familiar categories, established routines, and inherited metrics shaped by prior experience, which function as cognitive filters through which new developments are assessed. Under stable conditions, this form of cognitive economization is efficient. It becomes problematic, however, when a technology alters what those inherited categories and metrics actually capture. In such cases, indicators that once reliably tracked key organizational dimensions can appear to improve precisely as their representativeness declines. These interpretive habits are stabilized by attention structures, including dedicated channels, established agendas, and entrenched routines, which make them resistant to revision even when misalignment between indicators and underlying realities has the potential to generate significant harm. This structural stickiness is a primary source of organizational inertia. This thesis utilizes the diffusion of Generative Artificial Intelligence (GenAI) as a contemporary case study to investigate the interaction between technological novelty and organizational inertia through an Attention-Based View. The overarching research question guiding this thesis is: How do generative technologies reshape the relationship between observable indicators and the underlying organizational capabilities they are meant to track? Building on the Attention-Based View, it develops the Visibility Trap framework: a self-reinforcing mechanism in which established attention structures privilege impacts that are legible and strong within existing evaluation channels, while slower or harder-to-code consequences remain peripheral and therefore under-managed. The framework is applied across two phases of GenAI diffusion. In the early governance phase, a case study of the Italian Data Protection Authority's temporary ban on ChatGPT illustrates how regulatory action can register as a procedural and symbolic success, eliciting provider adjustments, financial penalties, and international signaling, while producing legitimacy erosion and limited behavioral change. In the second phase, namely organizational deployment in knowledge work, a theory-building study informed by exploratory interviews demonstrates that productivity gains and apparently improved access to organizational memory can coexist with shallow learning, growing dependence on GenAI for baseline task performance, and the rise of pseudo-memory – synthetic precedents that appear authoritative but are weakly grounded in the organization's past. Across both domains, the findings indicate a consistent pattern: GenAI amplifies the salience of historically established success signals, while weakening their value as proxies for underlying capability, efficacy, and epistemic integrity. The thesis contributes to research on organizational myopia by grounding incubation and success traps in attention structures and evaluative architectures, offering a portable mechanism for explaining why visible improvements can systematically delay correction of latent erosion. It concludes that effective GenAI integration requires organizational redesign that actively examines whether prevailing measures continue to track the capacities and qualities they were historically intended to protect.