Bias

In machine learning, bias refers to a systematic deviation: a pattern by which an AI system consistently shifts certain outcomes in one direction. The term is neutral at first — it describes a tendency, not a judgment; a bias is not inherently bad. In the narrower statistical sense, bias denotes the systematic component of model error (the bias-variance trade-off) or the bias term of a neuron. In the broader sense, it arises in several ways: through (pre)judgments of people that flow into the training data; through data that only partially reflects reality (for example, when certain groups are underrepresented); or through decisions made during algorithm design. Whether a bias is a flaw depends on whether the deviation is factually justified — a difference is not automatically an error. If it rests on a genuinely relevant factor, the model accurately reflects the world. If, however, it rests on features irrelevant to the task, on mere proxies for those features, or on measurement and sampling errors, the model is distorted — it measures something other than what it is supposed to measure. The useful question is therefore not whether a difference exists, but whether it is relevant and justified for the task at hand. The tricky part: unlike an individual human judgment, a bias in an automated system becomes a reproducible, scalable pattern — for better or for worse.