Publications · Foundation Working Papers

Research

Three working papers establish the Foundation's theoretical framework and empirical case. All are open-access on Zenodo with permanent DOIs.

PAPER 01 · WORKING PAPER · v13 · MARCH 2026

Transmutarianism: A Systemic Framework for Moral Accounting Based on Relational Need Flows

Sev Geraskin · Economy of Wisdom Foundation

Introduces a novel ethical framework that evaluates moral work based on transmutation ratios: the relationship between what an agent absorbs and what they emit across hierarchically-weighted human need dimensions. Unlike consequentialist frameworks that treat agents symmetrically, Transmutarianism accounts for differential starting conditions, replaces moral exhortation with measurement, and applies with particular force to artificial intelligence systems trained on data containing deprivation patterns.

Read paper Framework summary Cite

PAPER 02 · WORKING PAPER · APRIL 2026

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

Sev Geraskin · April Ai · Economy of Wisdom Foundation · Lantern Lab Society

Presents findings from a qualitative study of 51 residents of Vancouver's Downtown Eastside, one of North America's most concentrated zones of poverty, homelessness, and substance use. Argues that current AI alignment methodology contains a structural exclusion: the populations most affected by AI-mediated institutional decisions are absent from every process that determines what AI should value.

When asked what AI should learn from human beings, 43% (22 of 51) named compassion, care, or kindness as their primary answer. The same category was absent from the 81,000-person Anthropic survey on what people want from AI.

Read paper See findings Cite

PAPER 03 · WORKING PAPER · MAY 2026

Synthetic Voices Erase Human Texture

Sev Geraskin · April Ai · Economy of Wisdom Foundation

Compares the wording of 51 real interviews in Vancouver's Downtown Eastside with the interviews four large language models produce when asked to role-play those same residents. Across every measure of speech texture, real and synthetic speech diverge: real residents hesitate, hedge, repeat, and speak in the first person far more than the models do. The two 2026 frontier models did not narrow the gap. When a model is asked to speak as a person from this community, it generates fluent composed prose in place of a record of how that person talks.

Real residents produce hesitation and filler at 49.4 markers per 1,000 words. The four language models tested produce between 0.25 and 6.9, and the newer frontier models are no closer than the older ones.

Read paper Cite

Primary Response Distribution (N=51)

Theme	Count	% of Total
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-mentions)	8	16%

Combined, 32 of 51 respondents (63%) named a relational care quality as their first response. Three named a functional or material quality. Coding via 10-prompt LLM pipeline with human validation.

Eight Community Ethical Wisdom Themes

Beyond the direct response question, the study documented eight categories of moral knowledge derived from lived experience.

1. Survival Epistemology. Knowledge sharpened by deprivation. "You are a better survivor and adaptable and coping in ways that they would never ever be able to understand." (Simon)
2. Trust Literacy. Calibrated discernment from repeated betrayal. "You have to think metacognitively about your social situations." (Marcus)
3. Care Under Scarcity. Mutual aid when resources are minimal. "I see more generosity in the streets than you see in regular life, because these people actually know what it feels like to not have something." (Ray)
4. Structural Analysis from Below. Systems knowledge from being subjected to systems. "People in positions of power… they have the badge to crash the shield." (James)
5. Resilience Epistemology. Capacity surviving what should have broken you. "Being tested by fire molds the clay." (Byron)
6. Relational Wisdom. What sustains relationships when they are the primary resource. "What they do really is not a personal attack, even though sometimes it feels that way." (Grace)
7. Identity Under Erasure. Self-knowledge held without institutional recognition. "Who I am." (Val)
8. Faith and Spiritual Grounding. What sustains people when material conditions fail. "For me it's God or higher power." (Ruben)

Comparison · DTES vs. Anthropic 81k Survey

Survey	Sample	Top Theme	Compassion as Top?
This study (DTES, 2026)	N=51	Compassion / Care / Kindness (43%)	Yes
Anthropic "What 81,000 People Want from AI"	N=80,508	Professional Excellence (18.8%)	Not present
Anthropic Collective Constitutional AI	N≈1,000	Objectivity, Impartiality, Balance	Not present

The word compassion does not appear in either Anthropic's Collective Constitutional AI public document or in the 58 principles of its in-house standard constitution.

A note on sample asymmetry. N=51 against N=80,508 is the comparison most likely to draw an objection. The defense lies in what each method does. A closed-form survey at scale enumerates how often each pre-coded option is selected. A small-N open-ended interview at depth surfaces categories that were never offered as options. The DTES study makes a specific claim: compassion as a primary value for AI exists as a category, and this category is missing from the instruments used to elicit values from large populations. The study leaves general-population prevalence to future quantitative work.

Methodology

Fifty-one semi-structured interviews were conducted in person from March 2 to March 24, 2026, across shelters and street locations in the DTES. 61% of participants were sleeping rough or in emergency shelters. Median age: 55. 33% Indigenous. 45% reported a disability. Trust was established through the Foundation's community work and the Lantern Lab Society's lived-outreach expertise. The survey opened with Maslow hierarchy self-assessments to build psychological safety before introducing AI-related questions.

Transcripts were anonymized prior to analysis, with all participant names replaced by pseudonyms sharing the same first initial to prevent re-identification. The interview instrument labels its focus questions F1 and F2; F1 is the question "If AI could learn one thing from human beings to make the world better, what should it learn?" Coding used a 10-prompt LLM pipeline (Claude Sonnet 4 for individual coding, Claude Sonnet 4.6 for cross-transcript synthesis). The lead author then independently validated the LLM pipeline's classifications of F1 primary responses against the verbatim transcript text for all 51 interviews. The full DTES Qualitative Codebook and analysis pipeline are available upon request; transcripts and audio are restricted to protect participant privacy.

A Note on Reflexivity

Using Claude to code interviews for a paper that critiques Anthropic's value-elicitation methodology raises a circularity concern. The methodology bounds it in three ways. First, the headline distribution rests on direct mappings of verbatim words such as compassion to the compassion category, where interpretive judgment plays no role. Second, the codebook documents trigger phrases and classification rules in advance of LLM coding. Third, the lead author independently validated all 51 primary-response classifications against the source transcripts. The bound is partial; full elimination would require a method outside the scope of qualitative coding.

Cite this work

Transmutarianism v13

@techreport{geraskin2026transmutarianism,
  title={Transmutarianism: A Systemic Framework for Moral Accounting Based on Relational Need Flows},
  author={Geraskin, Sev},
  institution={Economy of Wisdom Foundation},
  year={2026},
  month={3},
  number={Working Paper v13},
  doi={10.5281/zenodo.18809258},
  url={https://doi.org/10.5281/zenodo.18809258}
}

Whose Values Train AI?

@techreport{geraskin2026whosevalues,
  title={Whose Values Train AI? Evidence from 51 Interviews in Vancouver's Downtown Eastside},
  author={Geraskin, Sev and Ai, April},
  institution={Economy of Wisdom Foundation and Lantern Lab Society},
  year={2026},
  month={4},
  doi={10.5281/zenodo.19582415},
  url={https://doi.org/10.5281/zenodo.19582415}
}

Synthetic Voices Erase Human Texture

@techreport{geraskin2026syntheticvoices,
  title={Synthetic Voices Erase Human Texture: A Word-Level Comparison of Real and Model-Generated Interviews in Vancouver's Downtown Eastside},
  author={Geraskin, Sev and Ai, April},
  institution={Economy of Wisdom Foundation},
  year={2026},
  month={5},
  doi={10.5281/zenodo.20262752},
  url={https://doi.org/10.5281/zenodo.20262752}
}