Lesson 1 of 4
Where harm comes from
The points in a project where bias and harm get in.
12 min 3-question quiz 3 guides to read next
Most harm from AI systems is not malicious. It creeps in through ordinary decisions at each stage of a project:
- The data reflects the past. The Adult census dataset describes incomes in 1994, including the pay gaps of that time. A model trained on it learns those gaps as if they were natural.
- Who is represented. Groups that are rare in the data get worse predictions, because the model has fewer examples to learn from.
- The label. "Was arrested" is not the same as "committed a crime"; "was hired" is not the same as "would do the job well". Choosing a convenient label can bake in someone else's judgement.
- The metric. A single overall accuracy figure can hide a model that works well for most people and badly for some.
- How it is used. A score meant to support a decision often ends up making it, without anyone reviewing the cases where it is wrong.
Removing a sensitive column such as sex or race does not remove the problem: other features (occupation, relationship status, even postcode) often carry the same information.
The following lessons turn these risks into concrete checks.
Check your understanding
3 questions · pass with 2 correct
Enrol for free to save your progress, unlock every lesson and earn a certificate.
Sign in to enrolFurther reading
Guides that go deeper on this lesson.
-
Understanding Bias in AI Systems
Where bias in AI comes from — data, labels, design and deployment — and why removing sensitive attributes isn't enough.
2 min read
-
What Is Responsible AI?
The principles behind responsible AI — fairness, transparency, privacy, safety, accountability — and how to turn them into practice.
2 min read
-
Human Oversight of AI Decisions
Designing human review that actually works: when it's needed, how to avoid rubber-stamping, and how to handle appeals.
2 min read