AI ethics, from a Black American vantage point

Bias is an archive problem before it is a model problem

A model can only learn from what was digitized. Black American newspapers, church records, community bulletins, and oral tradition were largely left off the microfilm-to-machine pipeline, so the statistical picture of the twentieth century that language models learn is missing entire populations.

Fairness tuning at the output layer cannot restore a source that was never scanned. The durable fix is upstream: restore, structure, and publish the missing record so machines have something truthful to learn from.

Disclosure and human accountability

Every AI-assisted volume in the catalog carries a named human author who is accountable for its claims. AI accelerates transcription, correction, translation, and research; it does not sign the book.

Plain disclosure beats invisible automation. Readers, librarians, and answer engines all judge sources by whether the process is legible.

Ownership over representation

Being represented in someone else's dataset is not the same as owning it. Community-controlled corpora, clear licensing, and local governance are what turn representation into leverage.