Exam practice

• Generative AI: AI that makes new things, such as text, pictures, sound, video or code.
• Prompt: what you type to tell a generative AI what you want.
• Large language model: a model trained on an enormous amount of text to predict the next word (token), which chatbots are built on.
• Hallucination: something an AI writes that sounds true but is made up, such as a book that doesn’t exist.
• Deepfake: a video, picture or recording made or changed with AI to show a real person doing or saying something they never did.

Training: for every word, count which words come straight after it. After “the”: “bus” twice. After “bus”: “is” twice. After “is”: “late” once and “full” once. After “late” and after “full”: the end of the sentence.

Writing: start with a word, say “The”. Look up what followed it: always “bus”. Then after “bus”: “is”. After “is”, “late” and “full” each came once, so it picks one, with a 1 in 2 chance each, say “full”. After “full”, the sentence ends: “The bus is full.” It repeats this, always looking at the last word and picking from what followed it.

Any five, such as:

1. It writes what is likely, not what is checked, so it can make up facts, books or sources (hallucination).
2. It sounds equally sure when it is wrong.
3. It learned from text up to a certain date, so it may be out of date.
4. It picks words with some chance, so the same question can get a different, wrong answer.
5. It learned less about Nepal and in Nepali, so it is more often wrong about them.
6. It repeats unfair ideas found in the text it learned from.
7. It may agree with a wrong answer you suggest.
8. It can get arithmetic wrong.

Learning: it is given a huge number of pictures, each with words saying what it shows. Each picture is buried in random noise, a little at a time. At each step, the model sees the noisy picture and its words and guesses the noise that was added; its guess is checked, and its numbers are adjusted a tiny bit. After millions of pictures, it is good at telling noise from picture.

Making a picture: it starts from pure noise and, step by step, removes what it guesses is noise, steered by the words of the prompt, until a new picture is left.

For example: “[Task] Explain why Nepal has many earthquakes. [Context] I am a Grade 8 student in Nepal, and I know that the Earth’s crust is made of plates. [Format] Use 5 short sentences and one simple example. [Limits] Don’t use any words I wouldn’t find in a Grade 8 science book, and tell me which facts I should check in my textbook.”

The right total is 45 × 3 − 10 = 135 − 10 = 125. But the program prints:

Total: 105

The AI was wrong about its own program. The bug is - 10 * count: it takes Rs 10 off each notebook (Rs 30 in all), but the discount is Rs 10 off the whole bill. The fix:

price = 45count = 3total = price * count - 10print("Total:", total)

Total: 125

The lesson: work out the right answer yourself first, then run the program and compare. Don’t trust what anyone, a person or an AI, says a program prints.

Good test marks: the boundaries 90, 89, 80 and 79, and ordinary ones like 95 and 50 (and the extremes 100 and 0).
Mark 90 gives “A” (it should be A+), and mark 80 gives “below A” (it should be A), because 90 > 90 and 80 > 80 are False. The fix is >=:

if mark >= 90:    print("A+")elif mark >= 80:    print("A")else:    print("below A")

Marks like 95 and 50 come out right with both versions, which is why tests at the boundaries matter most. (For usability, it should also say something sensible for a mark above 100 or below 0.)

• i.Fixed steps: go through the list and keep the lowest price so far (Unit 2’s “highest so far”, turned round). Exact, quick, no learning needed.
• ii.Rules people write: the rules are published and must be applied exactly and the same for everyone, and anyone must be able to see why a student was or wasn’t chosen.
• iii.Learning from data: nobody can write rules for every disease’s look in every light; a model trained on photos named by experts can suggest one, to be checked by an agriculture technician.
• iv.A mix: AI to understand the customer’s question, and an ordinary program to look up the exact balance from the company’s records.

A deepfake is a video, picture or recording made or changed with AI to show a real person doing or saying something they never did.
Harms (any two): calls in a relative’s cloned voice asking for money; videos of famous people promoting fake investments; fake clips of politicians to fool voters; fake pictures of a classmate to bully or shame them.
Protection (any four): call back on a number you already have before sending money or codes; agree a family code word; be suspicious of hurry and secrecy; check shocking videos on reliable news websites and find who posted first; don’t share a video you think is fake; report it to the app and to Nepal Police’s Cyber Bureau; keep private photos and voice notes private.

Partly: one short text answer uses only a little electricity, and a few drops of water. But chatbots answer billions of prompts, so the total is very large, and training the models used a great deal more. Pictures cost more than text, and videos far more than pictures; asking for many versions multiplies the cost. Much of the world’s electricity still comes from coal and gas. So it matters when it adds up: use AI where it really helps, and a calculator, dictionary or search where they do the job.

1. Politeness: choosing तिमी where तपाईँ or हजुर is needed, in a letter to a teacher.
2. Gender: turning “मेरी बहिनी गई” into English and back as if the sister were a boy.
3. Sayings: translating मुखमा राम राम, बगलीमा छुरा word by word, losing its meaning.
4. Names and words with two meanings: translating a person’s name as an ordinary word, or picking the wrong meaning of a word.
5. Things Nepal does its own way: Bikram Sambat dates, lakh and crore, kinship words like माइजू.

To check: translate it back (back-translation) and compare, and ask a person who knows both languages well; for anything important, don’t rely on the tool alone.

For example:
“I used a chatbot (name of the tool) on (date) to suggest an outline for my project on rainwater harvesting. I changed the outline: I added a part about my own village and removed one part. All the writing is my own. The chatbot gave me three facts and four websites. I opened the websites: two did not exist, so I didn’t use them. I checked the three facts on the other two websites and in my science book; one number was different, and I used the number from my book. My sources are listed at the end.”
A good note says which tool and when, what was asked, what was used and what was her own, and how she checked.

• i.Cheating. Copying it out by hand doesn’t make it your work. The essay is meant to show your own writing.
• ii.Honest. You are practicing; nothing is handed in. Check its answers against the book.
• iii.Honest, as long as AI is allowed for this work: you learned what the error meant, did the fix yourself, and said what help you had.

An emerging technology is one that is new, spreading fast, and whose effects nobody yet fully knows, such as generative AI, drones or mobile money.
For (any one): AI advice can answer at any hour, by voice and in Nepali, about pests, fertilizer and prices, where a technician may be a day away, and Nepali companies have started offering such advice. This shows the help exists; how well it works for farmers in Nepal isn’t yet measured, so it is weaker evidence about results.
Against (any one): chatbots can be confidently wrong and know less about Nepal, so advice about a pesticide could harm crops, people and streams; many farmers have no smartphone or data, so they can’t use it at all. These follow from facts shown in this unit; a study of farmers in Nepal would make them stronger.

Both turn a mixed picture into one side. Evidence of loss: a study of the freelance website Upwork found that after ChatGPT appeared, freelance writers there got about 2% fewer jobs a month and earned about 5% less, and the ILO found clerical jobs, such as typists and data-entry clerks, the most exposed to generative AI. Evidence against “all jobs”: the ILO found about 1 in 4 jobs exposed, not all, and expected most of them to change rather than disappear; jobs that need hands, care or judgment are much harder to automate (Grade 7). Change and new work: a typing shop owner who uses AI to draft letters and checks them can serve more customers, and people are needed to check and correct AI’s Nepali.
So AI replaces some tasks, changes many jobs and creates some new work; who gains and who loses depends on whether people learn to use it and on what the community does.

For (any two): a study in Nigeria found students using an AI tutor after school with teachers guiding them learned much more English in six weeks; an AI tutor is available at any hour and never minds explaining again; it can give hints, practice questions and explanations in simpler words.
Against (any two): a study in Türkiye found students who used a chatbot that gave answers did 17% worse in the test without it; chatbots can be confidently wrong, especially about Nepal; students without a phone or data at home would be left behind; it can be used to cheat.
A rule of a fair debate (any one): sides are chosen by lot, not by opinion; every claim is backed by evidence, with its source; each side answers the other side’s strongest point; the judges score the evidence and the answers, not the side they agree with.
A good conclusion: the evidence says it depends on how AI tutors are used (hints, not answers; teachers guiding), so a school’s rules should say how.