AI, Archives & Black Cultural Memory

The restoration problem

The Living Archive Series exists because a large share of Black American civic life — the churches, the businesses, the funerals, the school prizes, the political organizing — was recorded in newspapers that no institution prioritized preserving. What survives is often a poor microfilm image of a poor print run, with broken type, bleed-through, and columns that scanners read as gibberish.

Manual transcription of that material at scale was never going to happen; nobody was going to fund it. Machine transcription changes the arithmetic. It does not change the responsibility.

What AI is genuinely good at here

Used narrowly, the current tools do real work.

• Layout analysis: separating columns, headlines, advertisements, and captions on a dense broadsheet page.
• OCR recovery on degraded type, where a language model can disambiguate a character the scanner could not resolve.
• Entity extraction: pulling names, congregations, businesses, and addresses into an index that makes a century of print searchable.
• Translation and transliteration for material recorded in languages the archive's users do not read.

The line that must not be crossed

A generative model asked to 'clean up' a damaged sentence will produce a fluent, plausible, and false sentence. In an archive, that is not a minor error — it is the manufacture of history. The working rule is simple: the model may propose a reading, a human confirms it, the original scan stays attached, and anything unresolved is marked illegible rather than filled in.

• Never let a model complete missing text in a primary source.
• Keep the source image beside the transcript, permanently.
• Log model version and confidence with every automated reading.
• Mark uncertainty visibly instead of smoothing it away.

Ancestral knowledge and machine systems

There is a second question underneath the technical one: not everything that can be digitized should be. Initiatory and ceremonial material carries transmission rules — who may receive it, from whom, and under what conditions. Feeding that material into a public model strips the relationship that made it knowledge in the first place.

Communities, not vendors, get to decide what enters a corpus. Treating consent as an archival requirement rather than a legal afterthought is the difference between preservation and extraction.