Every new machine invites the same first question: what can it do? A calculator can compute. A car can carry five people at speed. An AI system can recognise a face, translate a sentence, or flag a patient's risk level in seconds. That question gets answered by testing the machine — run it, measure it, see what it does.
There is a second question, and testing never answers it: what should this machine do? A car that can reach 200 kilometres an hour on a school road can still be driven at 20. What a machine CAN do and what it SHOULD do are two separate questions. For AI, the second one is called ethics — not a brake on what technology can do, but a check on how it gets used.
AI Ethics names the values that should guide how an AI system is designed, built, and used. Ethics asks four things of a system: is it fair? Does it respect privacy? Does it help society? Can someone be held accountable for what it does? None of these four is a nice-to-have. AI systems now decide things that shape a student's marks, a family's loan, or a patient's health report. Passing the CAN question is no longer the whole test. This chapter is about the SHOULD question — the one testing alone never answers.
Picture an AI system built for a hospital. Its job: flag which patients need urgent care first. Test it the usual way. How often is it right? How fast does it answer? By those numbers, it does well — high accuracy, an instant response.
Here is the trap: surely a system that is fast and accurate has done its job. That is the technical bar, and this system clears it easily.
But suppose it clears that bar unevenly. It flags risk correctly for patients from some backgrounds. It misses risk more often for patients from other backgrounds. The average accuracy still looks high. The harm lands on whoever the system quietly failed.
That is the ethical bar, and it asks a different question: did the system treat everyone the way it should have? A system can pass the first bar and fail the second, on the very same set of numbers. Ethical AI is not only about being smart and efficient. It is about being responsible and trustworthy — and being right on average was never the same thing as being right for everyone.
In Class 6, you learned to pause before sharing anything online and ask one question: what am I sharing, and do I actually need to? Grade 8 takes that same pause and turns it into four sharper questions, run before any personal information leaves your hands.
Why is this information needed — does the app or service actually require it? Who will be able to see it, once it is shared? How long will it be stored — a one-time use, or a permanent record? And can it be deleted later, if you change your mind?
Picture the moment these questions get used. You are installing an app, and a screen appears asking for permission: Location, Contacts, Photos, Camera. It would be easy to tap Allow without a second thought — most people do. The same four-question pause compresses, in that moment, into two: why does this app need this, and who can see my data once I say yes?
Some information should never be shared at all, whatever the four questions' answers turn out to be — OTPs, bank details, passwords, PINs. No app legitimately needs those as a photo, a screenshot, or a message.
Here is the reframe that makes the whole habit make sense: privacy was never about hiding from the world — it is about keeping control of your own information. Deciding what leaves your hands, who gets it, and for how long: the four questions are simply how that control gets exercised, one permission screen at a time.
Everything so far has been about one class outnumbering another inside a training set. There is a second mechanism here, and it works differently.
A system trained mostly on data from one city, or one social group, may not perform well once it meets people from somewhere else. The problem here is not the ratio between classes. It is the scope of the whole training set — how much of the world it actually saw.
A facial recognition system is a clean case. Suppose it is trained mostly on images of people from one region. Suppose, too, that within that training set, every class is shown in perfectly equal numbers. Check the numbers, and they come back balanced — the dataset would be marked fair. That is the trap.
Balance inside a training set is not the same thing as covering the people a system will actually meet. A perfectly balanced set can still come almost entirely from one region. The system never learned how people look in other places — and it gets worse there, no matter how balanced its starting classes were.
Four steps bring this together, and each one takes real effort.
Use data that is varied and fair to everyone. This directly answers the regional coverage problem from the last section — it widens what the training set actually sees.
Test systems closely before deployment. Catch a gap before it reaches real users, not after.
Allow human review of important decisions. A system cannot check its own blind spots; a person must.
Check performance often. A system that was fair at launch can drift, quietly, as the group it serves changes.
None of this happens by itself, and none of it is one-time. Fairness takes careful design, steady testing, and responsible use — held together, not applied once and left alone.
A message about an exam being cancelled starts on one phone. Within minutes, it has reached thousands of others — forwarded, screenshotted, forwarded again — long before anyone has checked whether it is true. That is misinformation: information that is incorrect or misleading, shared with others. It has always travelled fast. What has changed is how convincingly it can now be made.
In Class 6, you learned to check three things about anything online that looks real but might not be: who posted it, where it came from, when it was posted. That check still matters. But AI tools have added a new problem underneath it: they can now generate the false content itself, not just mislabel something real. A realistic fake photograph. A news article no journalist wrote. A voice that sounds exactly like someone you know, saying something they never said. A video showing an event that never happened.
The reach of a false message goes further than the one rumour it started as. It can shift public opinion, sway an election, damage someone's career or reputation, weaken public trust, and damage social harmony — making people trust even the true things a little less.
Before you believe or forward the next message, four habits extend what Class 6 already taught you. Check the source of the information. Verify the facts on a website you actually trust. Read past the headline before deciding what a story says. And think critically — ask whether a message is designed to make you feel something strong enough that you forward it before you check it.
Technology can shape what a whole community believes. Whether it shapes that belief truthfully depends, every time, on whether the person holding the phone paused first.
A hospital's screen shows four words: "Patient Risk Level: High." A team of doctors is in the room. They do not act on the message straight away. They open the patient's report, check the data behind it, and one doctor says the finding should be checked again before anyone decides anything.
Notice what did not happen. Nobody said "the AI found High risk, so we are done here." The system's job ended at producing that message. Checking it, deciding what it means, acting on it — all of that stayed with people.
Here is the trap: once an AI system makes a recommendation, surely the responsibility for what happens next has shifted to the AI — after all, the AI did the thinking. It has not shifted. Accountability means humans stay responsible for what an AI system does, every time, without exception. A machine is never the answer to the question "who is responsible?"
Four questions test this, and they apply wherever AI is used — in a hospital, a school, a bank, a court. Who created the system? Who checks that it works correctly? Who fixes it when something goes wrong? Who answers for the harm, if harm happens? In the fields where the stakes are highest — healthcare, education, banking, law, public services — these four questions matter most. Each one needs a real person's name behind it, not a machine's.
AI can assist a decision. It can flag, suggest, and work through data faster than any person in the room. What it cannot do is hold the responsibility for what happens next. That stays with whoever created the system, checked it, and acted on what it said.
Five threads run through this chapter. All five answer one question: now that AI can do so much, what should it do, and who stays responsible when it acts?
AI ethics is the guide. It is not a brake on what is possible. It is a check on how AI gets used. Privacy is not about hiding from the world. It is about staying in control of what happens to your own information. Bias creeps in wherever the data behind a system is incomplete, unfair, or narrower than the world it will actually meet. Misinformation is not new, but AI tools can now generate it convincingly and spread it fast — which is exactly why checking before forwarding matters more, not less.
And through every one of these threads runs the same rule. The responsibility for what an AI system does never transfers to the system itself. It stays with the people who built it, who use it, and who could have checked it and did not.