Where does AI bias come from?

From six places, of which the training data is only one: historical bias present in past decisions the model learns to reproduce, representation gaps in who the data covers, measurement bias where the recorded label is a proxy for what you actually care about, aggregation bias from one model serving groups that differ, deployment bias from using a system for something it was not validated for, and feedback loops where predictions shape the data used to retrain.

Can an AI system be made fair?

Not in every sense simultaneously. Several reasonable definitions of fairness — equal error rates across groups, equal positive prediction rates, and predictions that mean the same thing for everyone — have been shown to be mathematically incompatible except in degenerate cases. So choosing a fairness criterion is a value judgement about which harm matters most in your context, and it cannot be delegated to a technical team or a library.

Does removing sensitive attributes prevent discrimination?

No, and it usually makes things worse. Other features act as proxies — postcode, name, education, purchasing patterns — so the model reconstructs the attribute you removed. Worse, once you no longer record it you cannot measure whether your system discriminates, which is why measuring by group requires collecting group data.

Can complex models be explained?

Partially. Methods that attribute a prediction to input contributions give you a plausible account of which features mattered, which is useful for debugging and for meeting a duty to give reasons. They are approximations of the model's behaviour rather than descriptions of its reasoning, they can disagree with each other, and for genuinely consequential decisions a model that is interpretable by construction is usually the better choice.

Who is responsible when an AI system causes harm?

Legally this is still settling, and practically the answer must be a named person or role inside the deploying organisation. Diffusing responsibility across a vendor, a model, a dataset and an operator is how harm goes unaddressed, so responsible practice means naming who owns the outcome before deployment rather than after an incident.

What does the EU AI Act require?

It tiers obligations by risk. Some uses are prohibited outright, including social scoring and certain biometric categorisation. High-risk uses — employment, credit, education access, essential services, law enforcement — require risk assessment, data governance, technical documentation, logging, human oversight, accuracy and robustness measures, and conformity assessment. Limited-risk systems mainly owe transparency, such as disclosing that a user is interacting with a machine.

AI Ethics: the questions that decide whether a system should exist

This is not the course where you are told to be careful. It is the course where the specific, documented failures are laid out, along with why several of them cannot be engineered away and have to be decided instead.

What this course sets out to do: give you the concepts and the vocabulary to raise the right objection at the right moment, and to answer one competently when it is raised to you.

What this course does not do: provide a compliance checklist that absolves you. Those exist and they are not a substitute for judgement.

What you are about to discover

Course contents

#	Lesson	Main goal	Time
1	Where bias comes from	Six sources, and why removing attributes fails	9 min
2	Fairness cannot be satisfied fully	Incompatible definitions, and choosing between them	9 min
3	Explainability and its limits	What attribution methods deliver, and what they do not	8 min
4	Privacy, consent and data	Training data, memorisation, inference, and the law	8 min
5	Accountability and regulation	Who owns the outcome, and what is now required	9 min
6	Recap and FAQ	Synthesis, a review checklist, and 12 questions	6 min
7	Quiz and attestation	Validate what you learned with 5 corrected questions	3 min

Is this course for you?

You build these systems and want to know what to check before shipping.
You decide whether to buy or deploy one and need to ask better questions.
You work in legal, risk or compliance and want the technical reality behind the terminology.
You are affected by these systems and want to understand what recourse exists.

No technical prerequisites, though Machine Learning makes lesson 1 richer.

What you will be able to do at the end

Name the six sources of bias and identify which apply to a given system.
Explain why fairness definitions conflict and what that means for your project.
Say what explainability methods actually provide, and when to prefer a simpler model.
Discuss privacy in terms of consent, memorisation and inference rather than only storage.
Identify who should be accountable and what makes a decision contestable.
Summarise what regulation requires for a high-risk system.

Estimated time

Around 45 to 55 minutes of reading.

Prerequisites and next steps

Prerequisites: none. Machine Learning and MLOps add depth.

Natural continuation:

MLOps — the practices that make auditability possible
Generative AI — copyright, synthetic media and consent
Computer Vision — facial recognition specifically

Frequent questions, answered in one line

Is this just compliance paperwork?

Some of it is documentation, and the substance is not. A model with a thirty-point error gap between demographic groups is broken in a way no paperwork fixes, and the documentation exists because without it nobody discovers the gap.

Should some systems simply not be built?

Yes, and saying so is part of competent practice rather than obstruction. Some regulation now prohibits specific uses outright, and beyond those, a system whose error profile falls hardest on people least able to contest it is a design problem rather than a bug awaiting a patch.

Is a human reviewing the output enough?

Only if the human has genuine authority, time and information to disagree. Review that consists of confirming what the model already said adds accountability theatre rather than oversight, and this is a well-documented pattern rather than a hypothetical.

Does any of this apply to a small internal tool?

Proportionately. A ticket-routing model needs an owner and basic documentation. The obligations scale with what happens to someone when the system is wrong, which is the principle underlying every serious framework.

Want to go deeper on responsible practice?

The premium catalogue covers fairness metrics, bias auditing, explainability tooling and governance frameworks hands-on, with a verifiable certificate after a 40-question examination. Included in every paid plan.

Ready? Start with lesson 1 →

What you are about to discover​

Course contents​

Is this course for you?​

What you will be able to do at the end​

Estimated time​

Prerequisites and next steps​

Frequent questions, answered in one line​