Steps Towards Value-Aligned Systems

Individual AI systems aren't the only problem: While much attention focuses on fixing biased algorithms or unfair AI models, the paper argues that value misalignment often stems from how these systems interact within larger organizational and social contexts—meaning we need to look beyond the algorithm itself to understand and fix real-world harms.

A "systems-level approach" is necessary: Rather than evaluating whether a single AI tool is fair or safe in isolation, the authors advocate for assessing entire sociotechnical systems (the combination of technology, people, processes, and institutions) to catch misalignments that emerge from how different components interact with each other.

Fairness problems illustrate the broader issue: The paper uses fairness violations as a concrete example, suggesting that many fairness failures aren't simply due to bad algorithms but rather result from how systems are designed, deployed, and governed across multiple organizational levels and decision points.

Current research focus is incomplete: The existing literature's emphasis on individual artifacts means we're missing critical sources of value misalignment that occur at the system level, potentially leaving important problems unaddressed even when individual components appear to work correctly.

New questions need answering: The paper identifies that researchers and practitioners need to develop new frameworks and ask different questions about how to ensure entire AI-equipped systems remain aligned with societal values, not just whether individual tools meet technical standards.