Abstract
STS Project: Genealogy, AI, and Actor-Networks
This thesis looks at how digital genealogy platforms encode the structural silences of American
slavery into their matching logic. Using Bruno Latour’s Actor-Network Theory (ANT) and Pinch
and Bijker’s Social Construction of Technology (SCOT), I argue that tools like Ancestry.com and
FamilySearch were built around documentary conventions that work well for European-origin
families with consistent surname records across generations, and that same design consistently
fails descendants of enslaved people whose ancestors show up in plantation ledgers rather than
named census entries. My own family is the case study. Joseph Fossett, who was enslaved at
Monticello and worked as Jefferson’s blacksmith, and Edith Hern Fossett, who ran Jefferson’s
kitchen, are my direct ancestors. They are hard to find in these platforms not because the records
do not exist but because the platforms were never designed to read them. The paper proposes
Reparative Engineering as a response: a design practice that is honest about the politics baked into
its data, surfaces gaps instead of hiding them, and treats oral history and community knowledge as
real evidence rather than footnotes.
Technical Project: Genealogy Record Auditor
The Genealogy Record Auditor is a single-page web application built with Vite, React, and
TypeScript on the frontend, backed by a Node.js/Express proxy server that routes requests to the
Anthropic Claude API. A researcher pastes in a primary source text (a census entry, plantation
record, church register, compiled family narrative) and the tool runs an AI analysis flagging four
categories of archival problem that come up constantly in records of enslaved people: missing or
ambiguous names, date gaps and inconsistencies, lineage gaps, and entries that contradict each
other. The results come back as structured JSON and render as a labeled breakdown with a plain-
language summary and a research roadmap of up to six next steps, each one naming a specific
record type, time period, and archive. There is also a personal context field where you enter your
own name and what you already know about your family, which causes the model to give you
directions specific to your lineage instead of generic advice. The proxy architecture exists to keep
the API key out of the browser. Current limitations around independent record retrieval and session
persistence are noted as future work.
How They Relate
These are not two separate projects that happen to share a theme. The thesis makes an argument
and the tool is built to demonstrate it. The argument is that current genealogy platforms fail Black
descendants by design, not by accident: the logic embedded in their algorithms was shaped by a
social order that withheld the conditions of documentary personhood from the people it enslaved,
and that logic did not disappear when the records got digitized. Without something concrete to
point to, that stays an abstraction. The Genealogy Record Auditor is what makes it concrete.
The tool puts Reparative Engineering into practice. Commercial platforms penalize families whose
ancestors were denied surnames by requiring surname continuity to return high-confidence results.
The Auditor treats that absence as something to investigate rather than a reason to return nothing.
Where other tools hide their uncertainty behind a confidence score, this one names the gaps and
tells you what to do with them. The evidentiary standard shifts to match the actual conditions of
the record rather than the assumptions of a design built for someone else.
The tool’s limitations also prove the thesis’s point, which I think is worth saying directly. The
Auditor only works with what you give it. It cannot pull records on its own, cannot account for
documents that were never digitized, and struggles when an input is basically just a name and a
dollar figure, which is what a lot of antebellum inventory entries actually look like. Those are not
bugs I forgot to fix. They are the technical face of a political problem that a single prototype was
never going to solve. The thesis names the problem. The tool takes a step toward it and then shows
you exactly where the work still needs to go.
