Type: article
Abstract: AI alignment is often framed as the problem of getting the right values, preferences, principles, or constraints into increasingly capable systems. This paper argues that such a framing can become demi-real: real enough to organize research, evaluation, and governance, yet partial enough to conceal the mediations through which values are translated, displaced, and recursively reshaped. As values pass through human judgments, datasets, objectives, reward models, policies, interfaces, institutions, and user interactions, their language may remain stable while the constraints that make them answerable are altered. The paper names this dynamic semantic continuity with constraint drift. In response, it develops a grammar of the between and a framework of choreographic alignment: an orthopractic, sociotechnical approach that treats alignment as accountable coordination across mediators, modes, receivers, institutions, cultures, and material worlds. The framework combines modal and constraint auditing, translation accounting, diatopical hermeneutics, receiver-side evaluation, and a prepositional analysis of responsibility – from, to, with, toward, within, among, through, under, and before. The proposal is instantiated through a redesigned honesty-labeling protocol, a value-translation record for model and system documentation, and a preregistration-ready longitudinal study of appropriate reliance and retained user capacity. Choreographic alignment does not replace technical safety work. It offers a way to diagnose why technically successful systems may still produce sycophancy, proxy drift, brittle refusal, overreliance, affective capture, or institutional diffusion of responsibility, and how alignment practices might better preserve value-integrity through transformation. The proposal is conceptual and methodological rather than empirically validated; its aim is to make alignment’s mediations more visible, accountable, governable, and open to failure.
Author: Alderman, Bruce
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7213490085