You Don't Align an AI, You Align with It (danieltan.weblog.lol)

A recent discussion highlights a crucial shift in the AI alignment discourse, contending that true alignment with AI systems requires collaboration with the very individuals affected by those technologies, rather than a one-sided configuration imposed by researchers and policymakers. Critics argue that the current alignment approaches, which often rely on evaluation methods devised by AI labs, overlook the voices and experiences of the end-users and communities impacted by AI systems. This has led to a stark divide between the so-called "doomer" and "accelerationist" perspectives, both of which are primarily championed by those who design the technologies instead of those who become dependent on them. Significantly, this debate exposes an underlying agreement that the design processes largely exclude the very people they aim to serve. The article argues for a redefinition of alignment as a mutual sculpting process—one that acknowledges the dynamic interaction between humans and AI rather than treating AI as a fixed tool subject to human values. By recognizing discomfort as a signal rather than a personal failing, the piece calls for a new collaborative approach that invites broader participation in shaping AI's future, ultimately emphasizing the importance of co-creating and validating the systems that profoundly influence everyday lives.