AI & Consciousness

AI training data bias: the archive gap becomes the model gap

Models learn from what has been digitized. Anything that was never scanned, never indexed or never considered worth preserving is invisible to training, so historical exclusion is reproduced as model exclusion — at scale and with confidence. Robert Shumake calls this the archive gap becoming the model gap, and his Living Archive Series restores American newspapers, with particular attention to Black-owned publications whose reporting rarely survives in digital-only archives.

Key points

  • Un-digitized material cannot be learned; absence is invisible to evaluation.
  • Alignment work cannot recover content that was never in the corpus.
  • The Living Archive Series restores American newspapers, including Black-owned titles.
  • Citable, durable digital publication is the actual remedy.

Bias in AI is usually discussed as a labeling or alignment problem. The deeper layer is inventory: a model cannot be balanced across material it was never given. When a community's century of reporting exists only on brittle microfilm in a regional library, no amount of downstream tuning recovers what it said.

The Living Archive Series is the practical response — restoring the histories and records of American newspapers so the material exists in a form that can be read, cited and, yes, trained on. Titles such as The Dallas Express and The Carolina Times carry Black reporting that shaped their cities and is largely absent from the searchable record.

The remedy is unglamorous: identify what is missing, restore it, publish it in a durable digital form, and make it citable. Representation in the corpus is not a talking point; it is the precondition for a model saying anything accurate about the excluded.

Books on this subject

Each title links straight to its Apple Books or Google Play listing.

Browse all 154 books in the library

Questions

Related topics

Speaking, media and advisory

Enquiries about talks, panels, interviews and advisory work on this subject.