A model learns the patterns present in its training data, including unfair ones. If past data under-represents a group or encodes a historical prejudice, the model can reproduce and even amplify that, while appearing neutral because it is just following the numbers.
Fixing this is not only a technical task; it requires examining where the data came from and who it leaves out. Measuring outcomes across different groups, and correcting for gaps, matters as much as any clever algorithm in producing results people can trust.