Algorithmic bias
When an AI system produces unfair or skewed results that disadvantage particular groups of people. It usually stems from patterns in the data a model learned from, and is one of the central ethical concerns in deploying AI.
A machine learning model learns from examples, and it absorbs whatever patterns those examples contain — including the unfair ones. If the data reflects historical inequities, or simply underrepresents some groups, the model can carry that forward, producing results that work less well, or less fairly, for the people it saw least of. This is algorithmic bias, and it can appear in anything from hiring tools to facial recognition to lending decisions.
The tricky part is that the bias is rarely deliberate and often invisible at first. A recruitment model trained on a company’s past hires may quietly learn to favour the kinds of candidates it hired before; a vision system trained mostly on one group of faces may perform worse on others. Because the model is only reflecting its data, the problem hides in plain sight until someone measures the outcomes across different groups.
Addressing it is an active and important area of work: auditing systems for disparities, building more representative datasets, and treating fairness as something to be tested for rather than assumed. For a technology increasingly used in consequential decisions, it is one of the questions that matters most.