AI bias and training data: why models get Black history wrong

Where the gap comes from

Web-scale corpora inherit the biases of what was digitized first and cheapest. Large newspaper archives were scanned by institutions with budgets; small Black-owned papers frequently were not. Oral and initiatory traditions were written about by outsiders far more often than by practitioners. The result is a corpus where the confident, abundant voice is the secondary one.

• Under-digitization of community and Black-owned publications
• Secondary ethnography outweighing practitioner accounts
• Non-English and diacritic-heavy terms degraded by poor OCR
• Sacred material held deliberately offline, so absence reads as nonexistence

What models get wrong in practice

Ask a general model about Ifá, Òrìṣà practice, the Tamil Siddhar tradition, or Thai forest Buddhism and you will typically get a fluent summary blended from popular sources, with lineage details, ranks, and ritual sequence quietly wrong. Fluency masks the error; the answer sounds like an expert and is not one.

The fix is archival

You correct a corpus by adding to it. Restoring newspapers, publishing practitioner-authored texts, and marking them up so they are attributable does more for model accuracy than any system prompt. Consent matters too: some knowledge is community-held and should be documented as restricted rather than scraped.