Requirements Debt in AI-Enabled Perception Systems Development: An Industrial RE4AI Perspective

TL;DR

This paper empirically explores how requirements evolve and accumulate as debt in AI-enabled automotive perception systems, affecting safety, reliability, and compliance.

Contribution

It connects technical debt theory with RE4AI, identifying mechanisms of Requirements Debt through empirical interviews with industry experts.

Findings

Evolving functional requirements cause semantic drift and validation backlog.

Non-functional requirements shifts create assurance lag and compliance misalignment.

Requirements Debt propagates across data, models, and system artifacts, impacting safety and certification.

Abstract

AI integration in automotive perception systems shifts requirements from static specifications to continuously evolving entities shaped by data, models, and operating contexts. When such changes are not consistently documented, validated, and traced, they accumulate as Requirements Debt (ReD), an underexplored but critical subtype of technical debt. This study conceptualises and empirically investigates how evolving functional and non-functional requirements create and propagate ReD across the AI-enabled automotive perception system lifecycle. We conducted 16 semi-structured interviews with experts from 13 international automotive companies and 3 European research institutes, and analysed the data using thematic analysis. As one of the first empirical studies connecting technical debt theory with RE4AI, the work identifies key ReD mechanisms. Evolving functional requirements (e.g., algorithm updates, sensor fusion, architectural changes, real-time constraints) drive semantic drift, validation backlogs, and integration debt when verification lags behind rapid iteration. In parallel, evolving non-functional requirements (e.g., safety, cybersecurity, reliability, scalability, transparency, trustworthiness) create assurance lag, compliance misalignment, and transparency and reliability debt as standards and ethical expectations shift. These interacting mechanisms propagate ReD across data, models, and system artefacts, undermining auditability, reliability, and certification readiness in safety-critical perception systems.