Bias: data and design, not malice

Bias in AI is not about someone programming a model to be unfair. Instead, it creeps in through the data used for training: if a dataset over-represents one group or under-represents another, the model learns those skewed patterns. Even careful design choices—like which features to include or how to label examples—can amplify existing social biases. For instance, a hiring model trained on past successful hires might favour men if the historical data already reflected gender imbalance, even though no one typed a sexist rule.

To spot bias in exam questions, look for clues about data sources: is the training set balanced? Are labels subjective or drawn from biased human decisions? A quick mental check is to ask: 'Would this model work equally well for a different demographic?' If the answer is no, bias is likely. Another trick is to remember that bias can appear even when the model's accuracy is high—accuracy on a skewed dataset can hide unfairness. Also, watch for questions that confuse correlation with causation; a model might learn a spurious link (e.g., postcode and creditworthiness) that reflects historical discrimination.

To lock this in, think of a simple contrast: a model trained only on cat photos from one breed will fail to recognise others—not because it's malicious, but because the data was narrow. The same logic applies to human bias: the model mirrors what it saw. When you see a bias question, first check the data, not the code.