Fair Machines? Bias and Ethics in Artificial Intelligence

Imagine a large company builds an AI tool to sort job applications. It studies ten years of previous hiring decisions to learn what a 'good' candidate looks like. A year later, the company discovers the tool is rating women lower than men for the same qualifications. Nobody programmed it to do this — yet the bias crept in anyway. How do you think that happened? Discuss your ideas before reading on.

Artificial intelligence systems learn by finding patterns in large collections of data called training data. The AI does not have opinions of its own; it simply reflects whatever patterns exist in that data. This is powerful — but it creates a serious ethical problem. When the training data contains historical human prejudices, the AI learns and repeats those prejudices. This is called algorithmic bias — a systematic, unfair skew in an AI's outputs that disadvantages certain groups of people. The word 'algorithmic' just means the bias comes from the rules and calculations the AI uses. There are three main routes through which bias enters AI systems: 1. Biased training data. If the data over-represents one group and under-represents another, the AI treats the two groups differently. The Amazon hiring tool from the starter had been trained on ten years of CVs from a male-dominated industry, so it learned to prefer male candidates. 2. Biased labels. Humans often label data to teach an AI what is 'correct'. If the labellers hold unconscious prejudices, those prejudices get baked into the AI's rules. 3. Choosing the wrong measure of success. If engineers only measure whether an AI is accurate on average, they may miss that it works well for some groups but poorly for others. This is called measurement bias. Why does this raise ethical concerns? Ethics is the study of right and wrong — what we ought to do and why. When an AI makes consequential decisions — about who gets a loan, who is stopped by police, whether a medical scan shows disease — a biased system can cause real harm to real people, often those from groups that are already disadvantaged. Because AI scales up rapidly (one algorithm can make millions of decisions), biased AI can do far more harm far faster than one biased human. Fairness in AI does not have one simple definition. Researchers debate whether fairness means equal accuracy across groups, equal error rates, or equal outcomes. This disagreement itself is an ethical challenge that engineers, governments, and society must work through together.