Bias: a systematic pattern needs a specific explanation

Bias can mean a systematic distortion in data or predictions. It also has narrower technical meanings: an additive neural-network parameter, an estimator's statistical bias, or the assumptions that make some patterns easier to learn than others.

These meanings should not be collapsed. A model's bias parameter is not a measurement of social unfairness.

Look past the overall result

A speech recognizer may work well on the dominant accent while failing on another. Possible causes include unequal training coverage, poor labels, recording conditions and a mismatch between the measured target and the real task.

Measure the relevant slices, their sample sizes and error types. Equal overall accuracy does not imply equal false-positive rates, and different fairness criteria can require different tradeoffs.

The useful question is which mechanism produced the disparity and which change would improve the intended outcome. Removing an obvious group identifier is insufficient when other features still carry correlated information.