Navigating the Green AI Trade-off: A Qualitative Exploration of AI Practitioners’ Cognitive Thresholds and the Role of Model Cards

Abstract

The attainment of global sustainability targets has suffered a significant setback owing to the unprecedented growth in the adoption of artificial intelligence (AI) technologies both by the organization and the society at large. Earlier efforts to tame this threat have largely ignored the impact of human element in attainment of sustainable AI with preference for algorithm optimization and hardware efficiency. Green AI achievement is not immune to the multitude of daily AI model selection decisions made by AI practitioners and the continuous sustainability exclusion from these model selection criteria threatens an unavoidable surge in global carbon footprints. It is based on this background that the thesis seeks to answer two key questions, first, how do AI practitioners navigate the Green AI trade-off within their cognitive limits and secondly, how can Green AI model cards close the operationalization chasm that is the gap between awareness and action of Sustainable AI.

Thus, this thesis focuses on promotion of Green AI adoption AI practitioners by applying a dual synthesis of theory of bounded rationality and Diffusion of Innovation. The “how” of the problem is diagnosed by using the bounded rationality theory which provided explanation for the “satisficing” behaviors of AI practitioners due to information overload when making AI model selection decisions and the diffusion of innovation theory complemented the bounded rationality by providing the remedies for reversing the ugly trend through the redesigning of the decision making artefact that is the Green AI model card.

This study uses a qualitative approach by conducting a semi structured interview with eight purposively selected domain experts. First, the thesis established the current model selection practice among AI practitioners then proceeded to introduce the redesigned Green AI model card as a research stimulus to ascertain its effectiveness in bridging sustainability exclusion and finally the determination of impact of compatibility attribute in transforming sustainability into routine habits for practitioners. Reflexive thematic analysis of the data revealed three important findings to Green AI operationalization. First, practitioners “satisfice” without sustainability that is they select good enough AI models solely based on performance and accuracy related parameter. Second, contextualization and color coding of environmental metrics on the redesigned model cards enables practitioners to overcome their bounded awareness by triggering their system 1 thinking. Third, all the participants prioritize compatibility of Green AI model card to their existing workflow as a prerequisite for speedy adoption. The participants also inadvertently offered inductive themes which provide novel strategies for overcoming institutional barriers to sustainable AI adoptions, such recommendations include CSR Framing and updated regulations from policy makers to enforce compliance.