Safety

Bias (in AI)

Systematic errors in AI outputs that reflect prejudices present in training data, model design, or evaluation criteria, potentially leading to unfair outcomes.

Explained at five levels

Level 1

When the AI treats some people unfairly because it learned from examples that weren't fair to begin with.

Level 2

When AI makes unfair or prejudiced decisions because the data it learned from had biases — like if most examples showed only one group of people.

Level 3

Systematic errors in AI outputs that reflect prejudices present in training data, model design, or evaluation criteria, potentially leading to unfair outcomes.

Level 4

Unintended systematic distortion in model outputs arising from training data composition, annotation practices, or optimization objectives — manifesting as representational, allocational, or stereotyping harms.

Level 5

Statistical or societal bias propagated through the ML pipeline — from sampling bias and label bias in data collection through inductive bias in model architecture to evaluation bias in benchmarking — requiring intersectional fairness auditing.

Definitions are educational summaries. Terminology can vary by source and context.

Sources