The hazard of fluency: a framework for governing AI liability in organizational decision-making
Year: 2026
Type: conference-paper
Abstract: As artificial intelligence transitions from numerical, discriminative algorithms to fluent Large Language Models (LLMs), human-machine interaction faces a profound paradigm shift.Historically, users exhibited algorithm aversion toward classical opaque systems.Nowadays, the conversational fluency of generative AI has triggered an opposing crisis: uncritical overreliance, frequently resulting in the CHOKE phenomenon (Certain Hallucinations Overriding Known Evidence).This paper argues that treating algorithmic trust as a linear metric linked to transparency to be maximized via Explainable AI (XAI) is fundamentally incomplete.Specifically, in highly conversational systems, unchecked trust becomes a severe organizational vulnerability.To address this, we propose a 3D Trust Calibration Framework, a multidisciplinary model that maps three dimensions of interaction: System Modality (Analytical vs Conversational), Decision Stakes (Low vs High), and Organisational AI Governance (Frictionless vs Structured).Drawing on Cognitive Psychology and Agency Theory, we deconstruct how in working environments conversational fluency of AI systems drives employees to engage in "unseen sub-delegation", treating the LLM as a shadow agent.This creates a severe governance hazard for the organization: while the employee minimizes their cognitive effort and acts as a superficial liability buffer, the organization unknowingly absorbs the reputational, legal, and operational consequences of algorithmic failures.Moving beyond purely technical paradigms of AI transparency, our theoretical framework establishes a new architectural ontology of trust and AI governance in modern organisations, providing a structural blueprint for designing accountable, friction-positive decision workflows.
Authors: Galera Zarco, Carlos, Zieni, Baraa
Institution: University College London
Language: English
Subfield: Safety Research
Field: Social Sciences
Domain: Social Sciences
Open Access status: green
OpenAlex ID: W7225020746