From Aspiration to Specification: Requirements Engineering for Inclusive, Human-Centered AI

Abstract

Responsible AI literature is abundant in ethical principles, yet poor at offering practical, useful methods. Ethical guidelines, fairness metrics, and governance mandates proliferate, but structural discrimination, exclusion, and harm persist across deployed AI systems. In this keynote, I argue that the recurring gap between high-level principles and real-world practice is, at its core, a requirements engineering concern. Diversity, inclusion, transparency, and accountability are treated as aspirations to be endorsed rather than requirements to be elicited, specified, validated, and evolved. Drawing on a multi-year research program at CSIRO (Australia's National Science Agency), the talk traces a single structural diagnosis across four empirical studies. I present a systematic literature review of diversity and inclusion in AI, an analysis of real-world AI incidents, an evaluation of AI transparency statements, and an industry co-design study with an employment marketplace. I argue that meeting the required criteria does not guarantee validation against the needs of those most exposed to AI-supported decisions and potential harms. The keynote also presents the essential components of an Inclusive AI Toolkit: a risk framework, a set of role- and harm-aware guidelines, an assessment question bank, an AI incident analyser, and a user-story template. This collection of resources is offered as an operational mechanism to reframe inclusion in AI not only as a preventative measure, an ethical safeguard, and a harm-reduction strategy but also as an engineering solution for the redistribution of power over who defines the needs, risks, safety and justice in AI.