Whose Imagination is Trustworthy? A Synthesis
The Tuskegee Syphilis Study began in 1932 as medical research and ended in 1972 as genocide. For forty years, U.S. Public Health Service physicians withheld penicillin treatment from 399 Black men—even after it became the standard cure in 1947. When Peter Buxton, a PHS social worker, raised ethical concerns in 1966, both his superiors and a blue-ribbon panel of physicians dismissed him. The study continued. When the scandal broke in 1972, African-American men's life expectancy dropped by up to 1.4 years due to destroyed medical trust that persists to this day.
The pattern repeats. In the 1990s, Arizona State University collected DNA from 400 Havasupai tribal members with promises of diabetes research. Instead, their blood was used for studies on schizophrenia, inbreeding, and tribal migration—topics sacred and taboo in Havasupai culture. The researchers saw DNA; the Havasupai saw violation of sacred material. It took litigation and a 2010 settlement to force the university to recognize what Indigenous knowledge holders knew immediately: that extracting genetic material without understanding cultural context causes spiritual harm no consent form anticipated.
What links Tuskegee, Havasupai, residential schools, and current AI development is epistemological: those with power to deploy technology cannot reliably imagine harms to those without power. Tuskegee doctors couldn't envision doing to their sisters what they did to Black men. ASU researchers couldn't fathom that "scientific progress" violated sacred worldviews. Residential school architects couldn't imagine cultural genocide because they framed assimilation as education. This isn't individual moral failure. It's structural blindness that emerges when one group designs systems affecting another.
British philosopher Miranda Fricker identified two types of epistemic injustice—silencing that directly affects marginalized testimony, and hermeneutical gaps where entire concepts for understanding harm don't exist in dominant frameworks. Before 1975, "sexual harassment" had no name; before 2015, Canada's Supreme Court hadn't recognized residential schools as genocide. The people experiencing these harms understood them long before dominant institutions had language to acknowledge them. This reveals a crucial asymmetry: those targeted by harm possess knowledge inaccessible to those causing it. This isn't opinion or perspective. It's epistemic privilege rooted in survival.
When disability advocates challenged AI hiring systems, they identified what technologists missed: algorithms can treat deviation from statistical norms as "outlier," constructing disability through technology itself. When Indigenous Elders expressed hesitancy about technology adoption, their caution wasn't technophobia—it was wisdom from experiencing extraction. Their ancestors' experiences had taught them that systems you don't control can extract resources, data, and culture. When residential school survivors consulted on a virtual reality recreation of Fort Alexander, they insisted the experience not be gamified, that viewers feel child-sized vulnerability, that their testimony be the core, not the technology. They understood harms developers couldn't.
The governance implication is radical: safety assessments conducted without those most at risk are epistemically incomplete. A blue-ribbon panel of physicians cleared Tuskegee; Black men had no voice. Tech companies assess AI safety; marginalized communities experience algorithmic harm. This isn't a fairness issue—it's a knowledge problem. Missing the people most vulnerable means missing the knowledge essential to accurate risk assessment.
Yet here's the paradox: this limitation isn't unique to AI or colonial-era medicine. It applies to any system where power holders assess impacts on the powerless. The solution isn't better intentions or more diverse hiring. It's structural: those who will experience worst outcomes must have veto power, not consultation. Consultation means "we listened." Veto means "you decide." The difference is sovereignty.
Indigenous communities in Canada have begun demanding Free, Prior and Informed Consent (FPIC) for development projects, not as obstruction but as foundation for sustainable co-governance. Some disability organizations have embedded lived experience into decision-making structures at parity with expert input. Australia's mental health sector developed "Lived Experience Governance Frameworks" that shift from "your input is valuable" to "your authority is structural." These aren't soft governance improvements. They're epistemological reorientation—recognizing that knowledge produced through suffering is as valid as knowledge produced through credentials.
The Haudenosaunee Seven Generations principle offers a temporal dimension. Rather than asking "is this profitable this quarter?" or "will this get me re-elected?", it asks "what will great-grandchildren's great-grandchildren think of this decision 140 years from now?" For AI, this reframes entirely. Today's deployment decisions don't just affect this decade. They propagate today's biases through systems that may shape humanity for a century. That weight requires epistemic humility—recognition that short-term optimization is inadequate.
So the answer to "Can an AI trained on colonial archives imagine colonial harms?" is not yes or no. It's structural. An AI that knows it cannot see what it cannot see—and therefore actively centers those who can—is fundamentally different from one claiming "safety" based on assessments by people who've never experienced being targeted. The work isn't building better AI. It's building governance that gives epistemic authority to those systematically denied it. Not because it's righteous, but because it's the only path to knowledge we can trust.
No comments:
Post a Comment