Whose Values Train AI? Evidence from 51 Interviews in Vancouver's Downtown Eastside

This paper presents findings from 51 in-person, semi-structured interviews conducted in Vancouver's Downtown Eastside (DTES) between March 2 and March 24, 2026. The study examines a foundational assumption in AI alignment research: that the human values used to train AI systems are representative of the populations most affected by them. When asked what AI should learn from human beings to make the world better, 22 of 51 respondents (43%) named compassion, care, or kindness as their primary answer. A further 6 (12%) named unconditional love. Combined with humanity and connection responses, 32 of 51 respondents (63%) identified a relational or caring quality as their first response. Functional or task-oriented answers were nearly absent. The finding stands in direct contrast to dominant AI values surveys, which report professional excellence, productivity, and reliability as top priorities among technologically literate populations. The paper argues that the alignment problem is not solely technical. It is also epistemic. Current alignment methodology, including RLHF, constitutional AI, and large-scale public surveys, collects values from populations that have been systematically screened for technological literacy, internet access, or survey panel membership. The DTES population is absent from these samples. This is a form of hermeneutical injustice at scale: an entire body of moral knowledge produced by the lived experience of institutional failure is unavailable to the processes that shape AI behavior because no mechanism exists to collect it. The study also documents a taxonomy of eight categories of community ethical wisdom derived from lived experience under scarcity, including survival epistemology, trust literacy, care under scarcity, structural analysis from below, resilience epistemology, relational wisdom, identity under erasure, and faith and spiritual grounding. The theoretical framework is Transmutarianism (Geraskin, 2026), which models relational dynamics as flows of fulfillment and deprivation between agents. The paper proposes a four-component field protocol for integrating system-affected populations into alignment pipelines as a distinct source of training data, with built-in ethical safeguards and participant protections. This work is part of a broader initiative by the Economy of Wisdom Foundation and Lantern Lab Society to connect emerging AI systems with communities historically excluded from technological design processes.