In machine learning, bias refers to the simplifying assumptions a model makes in order to learn general patterns from data, rather than memorizing every individual training example exactly. Some bias is a normal and necessary part of learning: without any simplifying assumptions, a model would struggle to generalize to new, unseen data at all.
The term also carries a separate, more consequential meaning: unwanted, systematic bias where a model's outputs unfairly favor or disadvantage particular groups, usually because the training data itself reflected historical inequities or unbalanced representation. In this second sense, bias is a major focus of responsible AI work; companies building large language models, including OpenAI with ChatGPT, invest specifically in identifying and reducing these unfair patterns so the model's responses don't systematically disadvantage particular groups of people.
