GLOSSARY
AI Bias
Systematic unfairness learned from data: models repeating and amplifying stereotypes about gender, race, language and culture at machine scale.
Models learn from human data, and human data contains human prejudice — so translation makes doctors male and nurses female, image generators whiten professions, and resume screeners downgrade names that skew minority. Bias is not an occasional bug but a statistical property of the training corpus, which is why “just fix it” doesn't work: the whole field responds with balanced datasets, evaluation suites for fairness, and RLHF preference data checked across demographics.
For users, practical defenses are concrete: ask the same question with swapped attributes and compare, demand sources for claims about groups of people, and remember that fluent confidence is not evidence of fairness. For builders, bias evaluations belong next to accuracy evals — shipping without them is shipping an untested bug at scale.