Direct discrimination is when somebody is disadvantaged because of a personal attribute (e.g. age, gender, race).

Indirect discrimination is when people with a certain personal attribute (e.g. age, gender, race) are disadvantaged even though the attribute is not explicitly considered in decisions.

Bias in the context of AI ethics looks at:

  • imbalance or tendency in data (input)
  • direct or indirect discrimination (output)

Examples of bias

  • A U.S. court trying to predict reoffenders were biased against black people.
  • Pedestrian detection miss rates on children.
  • Gender classification error rates based on race.

Causes of discrimination by AI

  • Risk is not anticipated, tested, or alleviated
    • “world bias”: world distribution problem
    • “representation bias”: not enough data is collected to accurately represent
    • “measurement bias”: wrong categorisation of people / wrong measurements
    • “algorithm bias”: wrong choice of algorithm
    • “evaluation bias”: wrong choice of evaluation metric or test set
  • Risk is obvious, problematic task
    • Ethics board does not flag a problem
    • Developer doesn’t oppose to build, doesn’t report, doesn’t blow whistle
    • Management makes decision to deploy