Every dataset reflects the world as seen from a particular angle, and that angle inevitably includes blind spots. Bias in AI training data isn't always the result of bad intentions - it often stems from practical constraints. Historical records over-represent certain populations. Web-scraped text skews toward English-speaking, internet-connected demographics. Medical datasets are dominated by studies conducted in wealthy countries. These imbalances translate directly into model behaviour. A hiring tool trained on historical decisions will replicate the biases embedded in those decisions. A facial recognition system trained mostly on lighter-skinned faces will perform worse on darker-skinned faces. A language model trained on internet text will absorb the stereotypes and prejudices found there. Addressing bias requires deliberate effort at multiple stages: auditing datasets for representativeness before training, testing models across demographic groups after training, and monitoring for disparate outcomes in deployment. There's no single fix - debiasing techniques can help but also introduce their own trade-offs, sometimes reducing overall accuracy or shifting the bias rather than eliminating it. The most honest approach is to be transparent about known limitations and invest in ongoing measurement rather than claiming the problem is solved.