The Ethics of AI: Bias, Fairness and Who Is Responsible

Artificial intelligence is making consequential decisions about people’s lives: whether to grant a loan, who gets interviewed for a job, whether a medical scan indicates cancer, whether a person leaving prison is likely to reoffend. These decisions were previously made by humans. Now they are increasingly made by algorithms trained on historical data — and that creates serious ethical questions that society is still working through.

AI ethics is not abstract philosophy. It is about real harms that have already happened and governance structures being built right now to prevent more of them. This guide explains the core issues: bias, fairness, transparency, and accountability.

The Bias Problem

AI models learn from data — and historical data reflects historical inequalities. A hiring algorithm trained on a company’s past successful hires will reproduce that company’s historical hiring patterns, including any historical discrimination. A facial recognition system trained primarily on white faces will be less accurate on darker-skinned faces, as documented in MIT research that found error rates up to 34% higher for darker-skinned women compared to lighter-skinned men.

Bias can enter AI at multiple stages: biased training data, biased problem framing, biased evaluation metrics, and biased deployment contexts. Removing bias requires deliberate effort at every stage — and even well-intentioned teams miss it because bias can be subtle and statistical.

Real-World Examples of AI Harm

The COMPAS recidivism algorithm, used in US courts to predict likelihood of reoffending, was shown by ProPublica in 2016 to be nearly twice as likely to falsely flag Black defendants as high-risk compared to white defendants. The algorithm was used in sentencing decisions affecting people’s freedom.

Amazon’s experimental hiring AI, built to screen CVs, was found in 2018 to systematically downgrade CVs from women — because it was trained on a decade of predominantly male successful hires. Amazon scrapped the tool.

Medical algorithms used to allocate healthcare resources in the US were found to systematically underallocate resources to Black patients compared to white patients with the same objective health needs, because the training proxy (healthcare spending) reflected historical disparities in access rather than health needs.

Fairness: Harder Than It Sounds

The mathematical concept of fairness turns out to be deeply complex. Researchers have identified over 20 distinct mathematical definitions of “fairness” — and many of them are mutually contradictory. You cannot simultaneously satisfy all of them in a single model. Which definition of fairness to optimise for is an ethical and political choice, not a purely technical one.

This means that “making AI fair” is not a problem engineering can solve alone. It requires genuine input from affected communities, ethicists, policymakers, and domain experts — not just data scientists.