AI In Practice: Functional Benchmarking and the Shaping of Societal Values

This thesis analyzes the nuances within human-computer interaction dynamics, specifically for artificial intelligence (AI). It focuses on the presence of benchmarking strategies in the evaluation processes of AI prior to deployment to human users. In particular, it explores the differences between technical benchmarking (engineering-focused evaluation processes) and functional benchmarking (real-life application assessment). It underscores the value of functional benchmarking in safeguarding individuals from the risks of potential AI hallucinations or biases in practical use. This is increasingly becoming relevant due to the deployment of AI technology in high-stakes domains like law, healthcare, and finance, where hallucinations and biases can have significant, irreversible societal implications. This thesis sheds a light on the lack of information and knowledge surrounding functional benchmarks and their place within AI assessment strategies for human-facing applications. This thesis explores the following research question: How can functional benchmarking serve as a lens to examine how AI systems reflect, sustain, or challenge societal values, and why has it not been widely adopted despite its potential to evaluate tangible real-world impact? The methods employed in this research study include two case studies - medicine and race and gender - and three interviews, analyzed with qualitative coding diagrams. Through these methodologies, this thesis asserts that functional benchmarking is pivotal in aligning the development of AI with current societal values and standards. In addition, this research outlines the barriers to functional benchmarking adoption, involving standardization challenges, reluctance to document mistakes, and broad conceptual barriers in defining fairness. Using qualitative analysis methods, this research emphasizes that functional benchmarking - when created collaboratively and transparently - can serve as an essential component in measuring model performance and promoting fairness, accountability, and societal value alignment. Fundamentally, this thesis positions functional benchmarking as an infrastructure for navigating the complex interplay between innovation and societal welfare in the age of intelligent machines.