CRAYON: Supporting the Elicitation of Pluralistic Normative Requirements for Autonomous Systems

Abstract

Normative requirements (NRs) capturing social, legal, ethical, empathetic, and cultural (SLEEC) expectations are increasingly critical for autonomous systems operating in humancentered environments. Governance frameworks such as the EU AI Act articulate these expectations as high-level principles, but translating them into concrete, capability-grounded requirements remains challenging in multidisciplinary settings. Existing work supports the analysis of normative rules once specified, yet offers limited methodological and tool support for the elicitation workflow itself. We introduce CRAYON, a tool-supported, humanin-the-loop methodology designed to support multidisciplinary stakeholders in eliciting normative requirements from high-level principles. CRAYON structures elicitation as an iterative process that links principles to system capabilities through proxies, rules, and concern-driven refinement, while maintaining explicit traceability. LLM-assisted recommenders generate candidate artifacts at each stage, and stakeholders retain responsibility for validation. We evaluate CRAYON with SLEEC experts on two real-world case studies and compare the resulting rule sets with a prior manual multidisciplinary elicitation. Results show that the CRAYON tool-supported workflow improves capabilitygrounded coverage and supports systematic refinement while preserving expert judgment and stakeholder creativity, nearly doubling the number of elicited rules (49 vs. 25) and achieving 86% stakeholder acceptance of generated candidates.