■ SAFE RISK ■ Education
No. The primary sources are elders, families and a community that decides what may be shared — and gaining that access requires standing in the community, not a research budget.
“Preserving a language that saved the world. This isn't about efficiency.”
Our AI replacement risk score — how we score jobs
The work spans archive and living memory. A historian in this field conducts oral history interviews in Diné bizaad with descendants and the dwindling number of surviving witnesses, cross-references Marine Corps unit records and declassified signals documents, catalogues photographs and personal effects, and builds exhibits and curricula for tribal schools and museums. There is a heavy public-facing component: veteran commemoration events, testimony for recognition efforts, advising documentary makers who arrive with an angle already chosen, and correcting the persistent mythology that turns a complex signals programme into a movie plot.
Digital tools have made large parts of this faster. Optical character recognition opens up typed military records, machine translation and speech recognition — now improving for low-resource Indigenous languages — assist transcription of interview tapes, and metadata systems make dispersed collections searchable. Language-revitalisation projects use the same technology to build teaching corpora. A model can draft an exhibit label or summarise a unit's operational history in minutes.
What it cannot do is earn the interview. Families decide who gets to record a grandfather's account; some material is culturally restricted and stays that way. Judging when a recollection has drifted, understanding kinship and clan context that determines who speaks for what, and negotiating data sovereignty over recordings — those are relational and political tasks. The field is also tiny, grant-funded and mission-driven rather than efficiency-driven, which is precisely why our score sits at 8. Consider a typical failure: a model transcribes a recorded interview, mangles a clan name, and that error propagates into an exhibit label, a school worksheet and eventually a Wikipedia citation. Catching that requires a person who knows the family, which is a qualification no dataset confers.
Automatability: our editorial assessment of current and near-term AI capability
Resilient, with urgency of a different kind. The clock here is biological, not technological: the last direct witnesses are gone or going, and the priority is capturing testimony from descendants before another generation passes. Transcription and archival search will keep getting cheaper through the 2030s, which helps small teams do more. The constraint on the field is funding and access, and neither is something automation resolves.
Partially, and it is improving. Speech recognition for low-resource languages has advanced enough to produce usable first-pass transcripts, which saves enormous time. But accuracy on elders' speech, dialect variation and archaic military-era vocabulary is unreliable, and errors in a historical record matter. The realistic workflow is machine draft, human verification by fluent speakers — which increases output rather than removing anyone.
It is a small, grant-dependent field usually attached to tribal museums, universities, the National Park Service or language programmes. Most practitioners combine it with teaching, archival work or broader Indigenous history scholarship. Jobs are scarce because funding is scarce, not because technology displaced anyone. Community connection and language ability matter more to hiring than credentials alone.
Time and money. The generation with direct memory is nearly gone, funding for tribal archives is chronically thin, and materials are scattered across federal collections, private hands and family homes. A secondary threat is bad automated history — models confidently generating plausible-sounding but false accounts that get repeated online, which historians then spend effort correcting.
Mostly by lowering the cost of the tedious parts: digitisation, OCR of typed records, searchable metadata, first-draft transcription, and 3D capture of artefacts. That lets tiny teams cover more ground. It also raises new governance questions — whether recordings should be publicly searchable at all, and who decides. Communities are increasingly asserting control over how their materials are digitised and indexed.