What is the difference between privacy and privacy protection?

Privacy is the thing at risk — the appropriate handling of personal data. Privacy protection is the principle you apply to keep it safe. The exam uses both words, and swapping them costs the mark.

Does explainable AI tell us who is responsible?

No. XAI makes it possible to understand why a judgement was made. Who answers for that judgement is accountability — a separate principle, and one the book says is still not settled between developer, operator and user.

Which principle answers which problem?

Bias → fairness (with transparency, because a bias nobody can see cannot be corrected). Surveillance and mass data collection → privacy protection. The black box → transparency, with XAI as the technology. Unclear responsibility → accountability.

Lesson 1-4

Ethical Issues with AI

Unit 1 — Information Technology and Society

Programming & Artificial Intelligence — Baccalaureate Year 2

Textbook pages pp. 26–32Published

Algorithmic bias, privacy, the black box and responsibility — and the four principles that answer them: fairness, transparency, privacy protection, accountability. Textbook pp. 26–32.

Learning objectives

01Explain what algorithmic bias is and describe its main causes in the data and in the way a system is built.
02Connect the four principles — fairness, transparency, privacy protection and accountability — to specific situations.
03Distinguish the problems (bias, privacy risk, the black box, unclear responsibility) from the principles that answer them.
04Judge an ethical situation involving AI, justify a rule for responsible use and say which parties carry responsibility.

Summary

The lesson starts where the last one stopped: AI now helps decide things about people — who is shortlisted for a job, who is identified by a camera, how personal data is used. Two facts follow from that. A system can be unfair if what it learned from was unfair, and it can be hard to find out why it decided anything at all. Those two sentences generate every issue in the lesson.

Algorithmic bias is bias in the AI's judgements caused by bias in the training data — a hiring system that rates one gender unfairly, a face-recognition system that misidentifies some groups more often. The causes come in two families: the data (too little of a group, or past discrimination recorded in it) and the build (a badly chosen variable, or a proxy variable that stands in for a protected attribute). The second family is why deleting gender from the data does not make a hiring AI fair: a postcode or a school name can carry the same information.

Two more problems and their names. Privacy is the handling and protection of personal data, and AI brings two new pressures on it: surveillance through face recognition, and mass collection of behaviour data. The black-box problem is an opaque decision process, which makes it hard to verify whether a result is correct; explainable AI (XAI) is the technology that tries to answer it. Responsibility is the open question of who answers when a judgement is wrong — developer, operator, user — and the book is explicit that it is not settled.

Against those four problems stand four principles: fairness (no unjust discrimination), transparency (showing the decision process clearly), privacy protection (handling personal information appropriately) and accountability (identifying who is answerable, and being able to hold them to it). Keeping problems and principles apart is half of every question in this lesson; the other half is knowing that XAI explains a decision while accountability decides who answers for it.

Key terms

Algorithmic bias: Bias in an AI's judgements caused by bias in its training data — for example a hiring system that rates one gender unfairly, or face recognition that misidentifies some groups more often.
Training data: The examples a system learns its patterns from before it is used. The less it represents reality, the more likely its outputs are biased.
Proxy variable: A variable that indirectly stands in for a protected attribute, so bias survives even after the attribute itself is removed from the data.
Privacy: The appropriate handling and protection of personal data. AI puts two new pressures on it: surveillance through face recognition, and mass collection of behaviour data.
Explainable AI (XAI): Technology that lets humans understand why an AI reached a particular judgement. It explains the decision; it does not decide who is answerable for it.
Black-box problem: An AI whose decision process is opaque, which makes it difficult to verify whether the result is correct or to detect an error or a bias.
Fairness: The principle that a system must not discriminate unjustly against a person or a group.
Accountability: The principle that the parties responsible for a system, its decisions and their effects are identified and can be held to account, each according to their role.

The lesson on video

Unit 1 · Part 4 — Ethical Issues with AI | Second Baccalaureate | English Schools

Watch on YouTube

Sample questions

01A company removes gender from the data its hiring AI learns from, and states that the system is now fair. Explain why that step alone does not remove bias, using the idea of a proxy variable.Apply
02A face-recognition system trained mostly on one group of faces is used at a border crossing to decide who is stopped. Put the issues in the order in which they arise: the cause in the data, the effect on the judgement, what it means for the person stopped, and why nobody can explain it — then name the principle each one calls for.Analyse
03A school wants an AI to rank applicants for its advanced class, and argues that a machine cannot be prejudiced because it has no feelings. Judge that argument, then propose one rule the school should adopt before using the system and justify it with two of the four principles.Evaluate

Answers are on the portal, graded there.

Questions students ask

Ethical Issues with AI

Open this lesson on the portal

Back to the unit All units