Online Communities as Collaborative Testbeds for Collective Alignment

Representing human values in machine learning (ML) applications has received increased attention as end-users of real-world deployments surface new failure modes. Many technical approaches for preference alignment capture stakeholder values indirectly by reflecting end-user agreement with ML outputs, but face multiple validity concerns related to what they optimize for. Participatory ML (PML) has emerged as an alternative mitigation strategy, but it faces parallel challenges that undermine both its feasibility at scale and stakeholder empowerment goals. We first argue that online communities are well-equipped to improve the quality of human feedback for preference alignment. To achieve these benefits, we position online communities as collaborative testbeds where members serve as active partners rather than passive subjects. We recommend three foundations for collaborative testbeds to navigate core challenges in PML: 1) community artifacts as legitimate proxies for consultation, 2) situated deliberation as feedback, and 3) collective meaning cascades as signals for evolving preferences, or criteria drift. By leveraging communities' existing capacities to articulate, contest, and carry forward nuanced value considerations, our position provides a critical direction for sustainably applying PML principles at scales meaningful for technical ML work.