Toward Engineer-Centered Ambiguity Detection in Regulated Requirements Engineering

Abstract

Natural language requirements remain the primary means of specifying software systems in regulated and safetycritical domains, yet ambiguity can lead to inconsistent interpretations, unclear verification criteria, certification delays, and safety risks. Although AI-based ambiguity detection has advanced considerably, many existing approaches still emphasize classification outcomes rather than the engineering decisions that follow detection. This position paper argues that, in regulated requirements engineering, ambiguity detection should be treated as one component of a broader human-centered decision-support process. We distinguish ambiguity detection from compliance impact assessment and refinement support, while showing how these tasks can be connected through an engineer-centered workflow. We propose a three-layer architecture comprising: (1) an interpretable evidence layer that combines linguistic indicators with transformer-based semantic analysis; (2) an engineer-configurable Bayesian decision layer that supports risksensitive thresholding; and (3) a generative AI assistance layer that provides regulation-aware refinement suggestions subject to engineer review. The paper articulates the theoretical grounding of this architecture, clarifies the role of human judgment and accountability, and outlines an empirical validation agenda for evaluating decision quality, trust calibration, workload, and regulatory grounding.