Deference by Design: Pluralistic Alignment Is an Interface Problem
Abstract
Pluralistic alignment is typically evaluated in terms of what models can represent: a range of reasonable answers, a specified perspective, or the preferences of a target population. But in deployment, it is the interface that often determines whether these capabilities reach the user. Even when models can express pluralism in theory, standard chatbot interfaces incentivize users to accept the model’s first fluent answer since contestation is cognitively taxing. We call this phenomenon deference by design, and trace its consequences for human autonomy and alignment. This paper makes three contributions. First, we reframe overreliance on AI defaults as a costrational response to AI systems and their interfaces rather than a user failure. Second, we argue that pluralistic alignment must be understood as a system-level property of models, interfaces, and users’ cognitive engagement budgets rather than one of the model alone. Third, we introduce Priori, an interface for human–AI complementarity that surfaces contestable choices in model responses as adjustable cards. Priori raises the abstraction level of oversight, making it cheaper for users to express values and revise hidden defaults. Together, these contributions motivate interfaces as a central design object for pluralistic alignment.