Bias.

In discussions of AI fairness, bias means systematic patterns that disadvantage people or misrepresent the world. The same word also has technical meanings, including an additive parameter in a model and assumptions built into a learning method.

Bias can enter through who is represented in the data, how labels are assigned, which objective is optimized, or where a model is deployed. A dataset can be large and still omit important situations. A model can achieve high average accuracy while performing much worse for a smaller group.

Examining bias involves choosing relevant populations, comparing error patterns, and understanding the consequences of those errors. There is no single fairness score that resolves every case: different definitions can conflict, and the appropriate tradeoffs depend on the task and its social context.

Removing sensitive attributes does not necessarily remove unfairness. Other features may act as proxies, and the labels or task design may already encode unequal treatment. Conversely, a neural-network “bias term” is not itself a claim about social prejudice.