AI Bias

Question 1

Skewed Sample

Answer 1

Occurs when the training data doesn’t fairly represent all groups or situations.

Ex: If an AI is mostly trained in English text from North America, it might not work well with content from other regions or cultures.

Question 2

Limited Features/Sample Size Disparity

Answer 2

Occurs when some types of data are much more common than others in training set.

Ex: AI has access to millions of formal English, but not many examples of casual or spoken english.

Question 3

Tainted Examples

Answer 3

Occurs when training data includes biased or incorrect info.

Ex: If AI learns from old texts with outdated views about certain groups of people, it may repeat those biased ideas.

Question 4

Proxy Bias

Answer 4

Occurs when seemingly natural info in the data is acually linked to sensitive topics.

Ex: AI might learn to associate certain universities (which may be linked to races or social class) with job stability, even if race or class aren’t mentioned in the data.

Question 5

Fairness Metric

Answer 5

Way to measure how fair an AI system is.

Question 6

Protected Class

Answer 6

A general category of people who share a charcacteristic that is legally protected against discrimination.

Question 7

Protected Feature

Answer 7

Specific instances or attributes within a protected class. These are the particular thins about a person that shouldn’t infuence the AI’s decisions unfairly.

Question 8

Pre-processing

Answer 8

Changing or fixing the training data

Question 9

In-processing

Answer 9

Charging how the AI learns during tarining to make it fair

Question 10

Post-Processing

Answer 10

Changing AI’s outputs after it has made a decision, to make the results fairer.