Algorithmic bias
What Is Algorithmic Bias?
Algorithmic bias is the systematic, repeatable skew in a computational system's outputs that disadvantages particular groups. In modern AI it originates mostly upstream of the model: in which documents were digitized, which were discarded during ingestion, and which communities were never recorded in machine-readable form at all. A model cannot report what its corpus never contained, so bias introduced at the archive stage survives every later correction.
Three places bias enters
Collection. The record itself is uneven — the Black press, community bulletins, church and fraternal records, and small-town papers were digitized late, partially, or never. Restoring 69 American newspapers for the Living Archive Series meant repeatedly finding that the only surviving run of a paper sat unscanned in one institution.
Processing. Optical character recognition performs worst on the material that most needs recovery, and automated correction trained on standard English rewrites names, dialect, and place references into something that never existed. The output looks clean and is wrong.
Deployment. A system evaluated only on aggregate accuracy hides failures concentrated in a minority slice of the data, because that slice is too small to move the headline number.
Why measurement is not repair
Fairness audits describe the symptom. They are worth running, but an audit cannot conjure a document that was never scanned. The intervention with the highest leverage is unglamorous: fund and execute digitization of under-represented archives, with human review by people who know the community the record came from.
Frequently asked
- What is algorithmic bias in simple terms?
- It is when a computer system consistently produces worse or distorted results for one group of people. In AI, that usually happens because the data the system learned from under-represented or misrepresented that group.
- Can algorithmic bias be eliminated?
- It can be substantially reduced but not eliminated by model work alone. Where source documents about a community do not exist in machine-readable form, the gap persists until those documents are recovered and corrected.
