Why this matters
An AI that learns from the past will repeat the past, including the parts we are trying to leave behind. A hiring model trained on years of decisions that favoured one group will keep favouring it, and it will do so in a tone of complete neutrality. That is the problem this lesson is about, and why four principles exist to hold it in check.
The idea
Algorithmic bias is bias in an AI's judgments that comes from bias in its training data. A hiring system that scores one gender down, or a face-recognition system that misidentifies some ethnic groups more often than others, are the standard examples.
It gets in more than one way. The data itself may be skewed, either because some groups are thinly represented or because past discrimination is recorded in it. But bias also enters through how the system is built: a badly chosen variable, a model design that amplifies a pattern, or a proxy variable, a harmless-looking field like a postcode that quietly stands in for something the model was never supposed to use.
AI also creates new privacy problems. Cameras in public spaces can identify and track individuals, and behaviour data gets collected at a scale no one consents to meaningfully.
Then there is the question of why. When a model's reasoning is a black box, nobody can check whether the answer was right, let alone fair. Explainable AI is the work of making that reasoning visible. And when an AI gets it wrong, who is at fault, the developer, the operator, or the user? That is genuinely unsettled, not a question with a filed-away answer.
Four principles are used to keep this in hand:
| Principle | What it asks for |
|---|---|
| Fairness | No unjust discrimination against a person or group |
| Transparency | The decision process and inner workings are shown clearly |
| Privacy protection | Personal information is handled properly |
| Accountability | Someone is answerable for the decision |
Picture it
flowchart TD
A[Biased training data] --> B[Model learns the bias]
C[Proxy variable or model design] --> B
B --> D[Unfair judgment]
D --> E{Can anyone explain it?}
E -- no --> F[Black box: the bias cannot be challenged]
E -- yes --> G[Explainable AI: it can be found and fixed]Worked example
An AI rejects someone's job application, and nobody at the company can say why.
Walk the four principles. Transparency fails first and most plainly: the process cannot be shown because it is not visible even internally. Fairness is now unknowable, which is not the same as fine; if the training data carried a past pattern of rejecting a particular group, this decision repeats it and no one would see. Accountability is where it gets practical: the applicant has a right to an answer, and "the model decided" is not one, so a person at the company has to own the outcome.
The fix is not to switch the AI off. It is to require that the model can report the factors behind a rejection, to test the outcomes across groups rather than trusting the model's neutrality, and to name a human who is answerable. That is explainable AI and accountability doing real work rather than sitting in a policy document.
Your turn
Try the practice questions for this lesson. Sort statements about bias and explainability into true and false, match each ethics principle to what it demands, and identify what explainable AI is actually for.
Take it further
A company is about to start using an AI system to screen job applicants.
- Investigate. Find one documented case where an AI system was accused of being unfair, and record your source. Which group did it affect, and what data might have caused it?
- Consider responsibility. When an AI makes an unfair decision, what share of the responsibility belongs to the developer, the company using it, and the operator?
- Decide. Name one rule the company should follow before it uses the system, and justify it using the four principles.
A rule that cannot be checked is not a rule. Ask how anyone would know whether yours was being followed.
Recap
- Algorithmic bias comes from biased data and from how the model is built, including proxy variables.
- An opaque model is a problem even when it is right, because nothing can be checked.
- Explainable AI makes the reasoning visible so it can be challenged.
- The four principles: fairness, transparency, privacy protection, accountability.
- Who is responsible for a wrong AI decision is still an open question.