Bias

Definitions

Systematic errors in an AI model's outputs that reflect prejudices in training data, algorithm design, or deployment context. An AI trained on biased data doesn't remove bias — it automates and scales it.
In plain English: When an AI system produces unfair or prejudiced results because the data it learned from contained those same prejudices — the AI inherits society's biases.
In statistics and machine learning, bias also refers to a systematic error in a model's predictions — underfitting due to overly simplistic assumptions. This technical bias (bias-variance tradeoff) is distinct from the ethical/social bias, though both cause models to be wrong.
Example: 'The model has high bias — it's too simple to capture the patterns in our data. We need more features or a more complex architecture.'
Source: statistical / ML technical

1960s: Statistical bias (systematic error) is well-established; cognitive bias enters psychology through Kahneman and Tversky in the 1970s
2016: ProPublica's investigation of COMPAS reveals racial bias in criminal recidivism algorithms, making AI bias front-page news
2020s: Bias auditing becomes standard practice; researchers develop frameworks for measuring and mitigating bias in language models