Bias in AI: Sources, Awareness and Mitigation

Hello students, welcome to Kwickprep. We think computers are neutral, because they only follow numbers. But an AI can still treat some people unfairly. Today we will see what AI bias looks like and where it comes from. Then we will play a hiring game, and learn how to reduce bias.

First, a new word. Bias means unfairly favouring one person or group over another. AI bias happens when an AI system gives results that are unfair to some groups of people. Most of the time nobody planned it, but the harm to people is still real.

Here is what AI bias can look like in real life. A hiring tool trained on past hires may keep rejecting women's applications. A face unlock system may fail more often for people with darker skin. A voice assistant may understand city accents but miss rural or regional accents. A loan app may give low scores to everyone from certain pin codes, even to people who always repay.

Let us look at a well known real case. A large technology company built an AI to screen job CVs. It learned from about ten years of past hiring, when most people hired were men. So it started giving lower scores to CVs with the word women's, as in women's chess club. The company found the problem and stopped using the tool.

Bias can enter at three stages. Data bias happens when the training data is unfair, old or missing some groups, like very few girls in a dataset. Design bias happens when builders make poor choices, like using pin code as an input, which can stand in for caste or income. Deployment bias happens when a system is used where it was not built for. An example is a city traffic app used in a village.

Data bias is the most common, so let us look closer. Historical bias means the data records unfair decisions from the past, and the AI copies them. Sampling bias means some groups are left out, for example photos collected only from big cities. Labelling bias means the people tagging the data bring their own opinions into the labels.

Now the other two sources. In design, builders may pick features or goals that are unfair, like judging students only by coaching fees paid. A team with only one kind of person may not notice problems others would see. In deployment, an AI may be used for a purpose it was never tested for. And if nobody checks the results after launch, bias can grow quietly.

Now an activity. Survival of the Best Fit is a free online game that shows how hiring AI becomes biased. You play a manager at a growing company. First, you choose applicants yourself from their CVs. Then, to go faster, you train an AI on your past choices, and you watch who it starts to reject.

Here is the lesson of the game as a flowchart. First, you hire people by hand, often in a hurry. Next, the AI learns from the choices you made. Then ask, were those past choices fair to every group? If yes, the AI has a better chance of staying fair. If no, the AI copies the bias and repeats it at a much larger scale.

After playing, discuss these questions. Why did the AI reject applicants from one group more often? Pause and think. Was the AI itself unfair, or was it the data it learned from? The answer is the data, because the AI only copies patterns. Finally, how could the company fix the problem?

So how do we reduce bias? First, collect diverse data, so every group is fairly represented. Second, remove unfair inputs and their stand-ins, because dropping one field alone is not enough. Third, test the results separately for each group, not just the average. Fourth, keep a human in the loop to review important decisions like jobs and loans. Fifth, build teams with different backgrounds, and keep checking the AI after launch.

Each source of bias has a matching fix. For data bias, collect diverse data and count how many examples each group has. For design bias, choose fair inputs and review every feature before training. For deployment bias, monitor how the AI is used and audit its results regularly.

Let us revise what we learned today. AI bias means an AI gives unfair results for some groups. Bias comes from three sources: data, design and deployment. Survival of the Best Fit showed that biased past choices create a biased AI. We reduce bias with diverse data, fair inputs, testing for each group, and human review. Next time you use an app, ask yourself, is it fair to everyone?