AI Ethics and Fairness

More than a technical question

For most of this subject we have asked how AI works. This lesson asks a different and harder question: should it work that way, and who is affected when it does? That is the territory of AI ethics. It is not a soft add-on to the engineering; it is part of doing the engineering responsibly. As AI systems increasingly decide who gets a loan, a job interview, a medical flag, or a longer prison recommendation, the stakes are real people's lives.

Ethics here is not about robots turning evil in films. The serious issues are quieter and already happening, and they grow directly out of the mechanisms you have learned, especially training data and bias.

Fairness is harder than it sounds

"Just make it fair" sounds obvious until you try to define fair. Suppose a system predicts which students might struggle, so a school can offer extra help. What does fairness require?

• Equal accuracy for every group?
• The same proportion of students flagged in each group?
• The same chance of being correctly flagged if a student really does need help?

Each of these is a reasonable definition of fairness, and here is the uncomfortable truth: researchers have proven that, in many situations, you cannot satisfy all of them at once. Improving one measure can worsen another. This means there is rarely a single "fair" setting an engineer can flip on. Instead, people must make a value-laden choice about which kind of fairness matters most for this situation, and be honest that others were traded away. Pretending fairness is a purely technical fix is itself a mistake.

The main ethical concerns

Bias and discrimination. Because models learn from data that reflects an unequal world, they can reproduce and amplify discrimination. A hiring tool trained on a company's past choices can learn to downrank certain groups; a face-recognition system trained mostly on some skin tones performs worse on others. The mechanism is covered in depth in Training Data and Bias in AI.

Transparency and explainability. Many AI systems, especially large neural networks, are hard to interpret. When a model denies you a loan, you deserve to know why, and a shrug of "the algorithm decided" is not good enough. Transparency means people can inspect, question and challenge how decisions are made.

Accountability. When a system causes harm, who answers for it? The temptation is to blame "the AI", but software has no responsibility; the people and companies who built and deployed it do. Clear lines of accountability stop harm from vanishing into a machine no one will own.

Privacy and consent. Modern models are trained on vast amounts of data, often scraped from the internet, including people's photos, writing and personal details, frequently without their knowledge or permission. Once your data trains a model, it is very hard to remove. Surveillance systems raise the stakes further, since powerful recognition tools can track people without their consent.

Power and access. Training frontier models costs millions and requires resources only a handful of companies and countries possess. That concentrates enormous influence in few hands and raises questions about who sets the rules, who benefits, and who is left out.

Environmental and labour costs. As covered in How Large Language Models Are Trained, training consumes significant energy and relies on low-paid human workers to label data and review disturbing content. Ethics includes the people and the planet behind the product, not just the user in front of it.

Why harm is usually accidental, and why that is no excuse

It would be easier if unfair AI were the work of villains. Usually it is not. Harm emerges from ordinary choices: a dataset that quietly under-represents some people, a target that optimises for clicks without considering wellbeing, a launch rushed before testing across groups. No one decided to discriminate; the system did it anyway.

This matters in two ways. First, it means good intentions do not protect you, so you must actively test for harm rather than assume its absence. Second, it means accidental harm is still harm. "We did not mean to" does not undo a wrongly rejected loan application or a wrongly flagged student. Responsibility is about outcomes, not just intentions.

What responsible practice looks like

Ethics is not only philosophy; it produces concrete habits. Strong teams:

1. Test across groups. Report accuracy and error rates separately for different genders, ages, skin tones and languages, not one flattering overall number.
2. Document the system. Publish what data it was trained on, what it is for, and where it is known to fail, so users and regulators can judge it.
3. Build in explanation. Design systems that can give reasons for high-stakes decisions, and give people a way to appeal.
4. Keep humans accountable. For decisions that affect lives, keep a person responsible in the loop rather than fully automating.
5. Respect consent and privacy. Collect only what is needed, get permission where possible, and protect what is stored.
6. Invite outside scrutiny. Let independent researchers and affected communities audit the system, because the builders are often blind to their own gaps.

Many of these ideas are now appearing in laws and regulations around the world, which try to set minimum standards for high-risk uses of AI.

Your part in this

You do not need to be an engineer to take part. As a user, you can ask sharp questions of any AI system: What was it trained on? Who might it fail? Who is accountable? Can I challenge its decision? As a future builder, you can treat fairness testing and transparency as part of the job, not an afterthought. The everyday side of this is explored in Using AI Safely and Responsibly, and if you want to see careers built around exactly these questions, look at Careers in AI. Ethics is not a brake on good technology; it is part of what makes technology actually good.