Whose Values Train AI? · April 2026 · DOI 10.5281/zenodo.19582415
The Evidence.
51 trauma-informed interviews in Vancouver's Downtown Eastside, March 2026. The populations most affected by AI-mediated decisions named what AI should learn from human beings. Their answers diverge sharply from the AI alignment field's existing data.
Headline finding
43% of 51 DTES residents named compassion, care, or kindness as the primary value AI should learn from human beings. The same category was absent from the 81,000-person Anthropic survey on what people want from AI, and absent from the 58 principles of Anthropic's in-house standard constitution.
Source: Geraskin & Ai (2026), Whose Values Train AI?Methodology
| Aspect | Detail |
|---|---|
| Population | 51 residents of Vancouver's Downtown Eastside (DTES). |
| Recruitment | In-person, through trust networks built by Economy of Wisdom Foundation and Lantern Lab Society. |
| Fieldwork window | March 2 to March 24, 2026. |
| Demographics | Median age 55. 84% male. 61% currently homeless or in shelters. 33% Indigenous. 45% self-reported disability. |
| Compensation | Cash at point of interview. Verbal informed consent. Pseudonym option offered to every participant. |
| Instrument | DTES Transmutarian Relational Flow Survey. Sequenced to build psychological safety: opens with Maslow and belonging items before introducing AI questions. |
| Coding | 10-prompt LLM pipeline. Claude Sonnet 4 for individual transcript coding. Claude Sonnet 4.6 for cross-transcript synthesis. Lead-author independent validation of all 51 F1 classifications against verbatim transcripts. |
| Ethics | Reviewed by Economy of Wisdom Foundation and Lantern Lab Society per community-based participatory research protocols. |
F1 Distribution: "What should AI learn from human beings?"
22 of 51 respondents (43%) named compassion, care, or kindness as their first answer. A further 6 (12%) named unconditional love. Combined, 28 of 51 (55%) identified a relational quality. Adding "humanity / connection," the relational care cluster reaches 32 of 51 (63%). Functional or task-oriented answers were nearly absent: only 3 of 51 (6%).
| Theme | n | % |
|---|---|---|
| Compassion / Care / Kindness | 22 | 43% |
| Love (unconditional) | 6 | 12% |
| Humanity / Connection | 4 | 8% |
| Human Sovereignty / Control | 3 | 6% |
| Reciprocity | 2 | 4% |
| Ecological Stewardship | 2 | 4% |
| Equality / Justice | 2 | 4% |
| Refusal / Skepticism | 2 | 4% |
| Other single-mention themes | 8 | 16% |
Comparison With Public AI-Values Surveys
Eight Community Ethical Wisdom Themes
The interview corpus organizes into eight categories of moral knowledge generated by hardship. Each entry below is the theme, a one-line definition, and a pseudonymized exemplar voice from the corpus.
| Theme | Definition | Exemplar |
|---|---|---|
| Survival Epistemology | Knowledge sharpened by deprivation. | Simon |
| Trust Literacy | Calibrated discernment from repeated betrayal. | Marcus |
| Care Under Scarcity | Mutual aid when resources are minimal. | Ray |
| Structural Analysis from Below | Systems knowledge from being subjected to systems. | Julian |
| Resilience Epistemology | Surviving what should have broken you. | Byron |
| Relational Wisdom | What sustains relationships when they are the primary resource. | Grace |
| Identity Under Erasure | Self-knowledge not depending on institutional recognition. | Val |
| Faith and Spiritual Grounding | What sustains when material conditions fail. | Ruben |
Participant Voices
Names are pseudonyms assigned by the project at the time of fieldwork; first initials are preserved.
Compassion. That whole compassion thing keeps coming up with AI for some reason. But it's true though. We need to learn that. I think yeah, a big heart to reflect a lot of its decisions from that big heart instead of, you know, having grudges.Sam
No matter how bad we treat it, it should always be loved. One thousand percent love, absolutely.Marcus
It must have compassion built in, hardware or software.Ben
Compassion, empathy, impartialness, non-judgmental. Like the things that I don't think it would know unless it was trained that way.Denise
I see more generosity in the streets or even inside of jail than you see in regular life, because these people actually know what it feels like to not have something.Ray
Being tested by fire molds the clay.Byron
See the Data
The 130-node relational network of the 51 respondents and their named organizations, systems, and contacts: network. Flows of fulfillment and deprivation through the same network, stratified by Maslow level: flows. A conceptual particle simulation of how the framework's flow primitives behave at population scale: simulator.