Towards a Framework for Operationalizing the Specification of Trustworthy AI Requirements

Requirements engineering is critical for trustworthy AI: The paper argues that building trustworthy AI systems requires structured approaches to specify requirements upfront, rather than treating trustworthiness as an afterthought. This means organizations need to systematically document what "trustworthy" means for their specific AI system before development begins.

AI systems need different specification approaches than traditional software: The authors highlight that machine learning systems behave differently from conventional deterministic software—they are "data-driven and non-deterministic"—meaning standard requirements engineering methods are insufficient. New frameworks like PerSpecML are needed to capture the unique challenges ML systems present.

Multiple stakeholder perspectives must be considered: The proposed framework emphasizes "perspective-based" methods to uncover concerns from different stakeholders. This suggests that trustworthiness requirements should incorporate input from diverse groups (users, affected communities, developers, ethicists) rather than being defined by technologists alone.

Traceability and structured documentation improve accountability: By combining AMDiRE's "artefact-centric" approach with stakeholder input, the framework aims to create clear documentation linking trustworthiness concerns to specific requirements and design decisions. This structured traceability helps organizations demonstrate how they addressed trustworthiness throughout development.

This remains an emerging challenge requiring further research: The authors conclude by noting "open challenges" still exist in operationalizing trustworthy AI requirements, indicating this is an evolving field where best practices are still being developed and refined.