Generative AI

The models you met in Grades 6 and 7 looked at something and gave it a label (Grade 7): a fruit sorter said “kagati”, a mango model said “ripe”. AI that makes new things, such as text, pictures, sound, video or computer code, is called generative AI (generate means to make or produce).

• A chatbot makes text: answers, stories, summaries, programs.
• An image generator makes pictures from words: type “a red bus on a mountain road in the monsoon”, and a picture appears that no camera ever took.
• Others make speech in a chosen voice, music, or short videos.

What you type to tell a generative AI what you want is called a prompt. What it gives back is its output.

Generative AI still learns from examples, as every model in Grades 6 and 7 did. But its examples are not labeled by people one by one. A chatbot learns from a huge amount of ordinary text, and the “right answer” for each example is already in the text: it is simply the next word. You will see how in the next section.

How did the phone’s keyboard know to suggest “birthday” after “Happy”? It had seen a great many sentences, and in them, “Happy” was very often followed by “birthday”. It counted.

You can build a model like that yourself. Take some text and, for every word in it, count which words come straight after it. In the sentence “Kathmandu is the capital of Nepal”, “is” is followed by “the”, and “the” by “capital”. Do this for every sentence, and you have a table of counts: that table is the model, and counting is its training. To write, the model starts with a word, looks up what followed that word in its text, picks one, and repeats with the word it just picked. Predicting what comes next, again and again, is how the model writes.

The model below learned from nine short sentences about Nepal, all of them true. Make it write.

From counting to a chatbot

Your model is tiny. A chatbot is built on the same idea, with three big differences.

1. It learns from an enormous amount of text. Not nine sentences, but a large part of everything people have written that its makers could collect: web pages, books, articles, computer programs, in many languages. More English than any other language, and far less Nepali.
2. It looks back much further. Your model looks only at the last word, so it forgets the start of the sentence. A chatbot looks at everything written so far, your question included, often many pages of it. So after “Mount Everest is the”, it knows the sentence is about Mount Everest, and “highest” is far more likely than “capital”.
3. It doesn’t store counts. No table could hold every possible stretch of words. Instead, a chatbot’s model is billions of numbers that are adjusted, during training, a tiny bit at a time, until they predict the next word well across all that text. The words are first split into small pieces called tokens, often a whole word, sometimes part of one. Grade 10 shows how.

A model like this, trained on a huge amount of text to predict the next word, is called a large language model. On its own it would just carry on any text you give it. To make it a helpful chatbot, it is trained further, on examples of good answers written by people, and on people’s ratings of which of its answers are better. That is why it answers in a friendly, helpful way. But underneath, it is still predicting the next word.

A chatbot looks back much further than your model, so it makes far fewer silly sentences. But it is built the same way: to write what is likely, not what is checked. So it can be wrong, in ways that are worth knowing.

• It makes things up. Asked for something it has no good answer to, it still writes the most likely-looking answer: a book that doesn’t exist, a quotation nobody said, a date that is wrong, a web address that leads nowhere (as Grade 7 warned about sources). Made-up output that sounds true is often called a hallucination. It is not lying: nothing in it checks what is true. It is doing what it was built to do, predicting likely words.
• It sounds sure either way. The same calm, confident sentences come out whether it is right or wrong. How sure it sounds tells you nothing.
• It may be out of date. It learned from text collected up to a certain date. Unless it searches the web as it answers, it knows nothing after that, and may not say so.
• It gives different answers to the same question. Like Roll the dice, chatbots pick words with some chance, so ask twice and the answer can change, sometimes from right to wrong.
• It knows less about Nepal. It learned what was written most. Far more is written in English about America than about Nepal, and far less still in Nepali, Maithili or Tamang. So its answers about Nepal’s history, laws, places and people are more often wrong, and its Nepali is weaker. This is Grade 7’s bias, one-sided examples, in a chatbot.
• It repeats people’s unfairness. It learned from what people wrote, including their unfair ideas about women, castes, religions and countries. Its makers try to train these out, but some remain.
• It agrees too easily. Tell it “I think the answer is 42” and it may agree, even when 42 is wrong.

An image generator is trained on a huge number of pictures, each with words that say what it shows: the captions and descriptions that people put with pictures on the internet. From them it learns which shapes and colors go with which words: what “bus”, “mountain road” and “monsoon” look like.

Many image generators are diffusion models, and they learn in a surprising way. In training, the computer takes a real picture and buries it in random noise (Unit 6’s word for whatever isn’t the signal you want), like the snow on an old TV, a little at a time, until nothing of the picture is left. At each step, the model is shown the noisy picture and its words, and must guess the noise that was added. Its guess is checked, and its numbers are adjusted a tiny bit. After millions of pictures, it is very good at guessing which part of a noisy picture is noise.

That is also why AI pictures go wrong in their own ways. The model knows what things look like, not how they work. Pictures made this way used to have hands with six fingers and signs with letters that spell nothing. Newer models make far fewer such mistakes, so, as Grade 7 said, don’t rely on spotting them Grade 7: Pictures that never happened. And a model trained mostly on pictures from other countries draws “a school in Nepal” or “a Nepali wedding” the way those pictures showed it, often wrongly.

A chatbot can only work from what you tell it, and a short, vague prompt gets a vague, general answer. Compare:

• “Tell me about water.” (About what, for whom, how long?)
• “Explain why water boils at a lower temperature at Namche Bazaar than in Kathmandu, for a Grade 8 student, in 5 short sentences, with one everyday example.”

The second says everything the chatbot needs. A good prompt usually has four parts:

1. The task: what to do. Explain, list, summarize, suggest, check, write.
2. The context: who it is for and what you already know. “For a Grade 8 student in Nepal.” “I understand loops, but not lists.”
3. The format: how the answer should look. “In 5 sentences.” “As a table with two columns.” “In simple Nepali.”
4. The limits: what to leave out or stick to. “Use only the facts in the paragraph below.” “Don’t give me the answer; give me a hint.”

A chatbot remembers the conversation so far, so you rarely need to start again. When the answer isn’t what you wanted, say what was wrong and ask again: “That is too long. Use simpler words, and give an example from a village.” Improving a prompt after a poor answer, step by step, is a normal part of using these tools well.

Grade 7 taught you to check claims by lateral reading, reading across other sites, and by finding the original source Grade 7: Who is behind it? Reading across. An answer from a chatbot needs exactly that checking, because it has no source at all: it is likely words. Ways to check:

1. Compare it with a source you trust: your textbook, a government website, an encyclopedia written and checked by people.
2. Check every name, number and date by searching for it yourself. These are where made-up facts hide.
3. Open every source it gives. If it names a book or a web page, find it, and see that it really says what the chatbot says it does.
4. Work out numbers yourself, with a calculator. A chatbot can get arithmetic wrong.
5. Ask yourself whether it makes sense, using what you already know. Something that surprises you most needs the most checking.

Chatbots write programs too, and quickly. Programmers use them every day. But a program from an AI is like a program from a classmate: it may have bugs, and you are the one who must test it. Check two things:

• Accuracy: does it give the right answer, for every input it should handle?
• Usability: can a person use it, and can a person read it? Does it ask clearly for what it needs, and say what its answers mean? Does it crash on a likely mistake, such as a word typed where a number was wanted? Would you understand it in a month, with its names and comments?

Here is a program a student got from a chatbot, asking for “a Python program that prints the average and the highest of these marks: 72, 85, 64, 90, 58”.

marks = [72, 85, 64, 90, 58]total = 0for i in range(1, len(marks)):    total = total + marks[i]average = total / len(marks)print("Average:", average) highest = 0for m in marks:    if m > highest:        highest = mprint("Highest:", highest)

Average: 59.4Highest: 90

It runs with no error. Is it right? Test it the way Unit 3 taught Grade 8: Testing an algorithm: with inputs whose answers you already know.

The second half, the highest mark, prints 90, which is right for these marks. But test it at the edges too. If every mark were below 0 (a list of temperatures in Jumla in winter, say), it would print 0, a number that isn’t in the list, because it starts from highest = 0. Starting from the first value, highest = marks[0], works for any list that isn’t empty. A bug that only appears for some inputs is the kind that a quick test misses.

Not every problem needs AI. A computer can solve a problem in three main ways, and part of a programmer’s job is to choose.

• Fixed steps: an algorithm that works the answer out the same way every time, like adding up a bill, or sorting marks in Unit 3. Use it when the method is known and exact.
• Rules people write: if-then rules that people decide, like a bus company’s fares, or grade boundaries. Use it when the rules are known and must be followed exactly, and when people must be able to see why the computer decided something.
• Learning from data: a model trained on examples, like the fruit sorter, voice typing, or a chatbot. Use it when nobody can write the rules (what does a face look like?), and when an answer that is sometimes wrong is acceptable, or can be checked.

Many real systems are a mix: a bank’s chat helper may use AI to understand a question typed in any words, then an ordinary program to look up the balance exactly, because a guessed balance would be a disaster.

Generative AI can copy a real person: their face in a video, their voice on a call. A video, photo or recording made or changed with AI to show a real person doing or saying something they never did is called a deepfake. With a few seconds of someone’s voice, from a video they posted, AI tools can make that voice say anything.

Deepfakes are used for jokes, and for films. They are also used to:

• cheat people of money: a call in a relative’s voice, “I’m in trouble abroad, send money now”, or a video of a famous person promoting a fake investment;
• fool voters: a politician seeming to say something they never said;
• bully and shame: a classmate’s face put onto a picture or video (Grade 7 Unit 9 covered cyberbullying; making such a picture is cruel, and can be a crime).

You can’t rely on spotting a deepfake by looking (Grade 7 said the same about AI pictures). Lips that don’t quite match the words, an odd blink, a voice with no breaths: these can be signs, but good fakes have none. So, as with any claim, check who and where instead of only what.

• A call or message asking for money, a code or a secret, in a hurry, even in a voice or face you know: hang up and call the person back on the number you already have for them, or ask someone who is with them.
• Agree a family code word: a word only your family knows, to ask for on any strange call “from” a relative.
• A shocking video of a famous person: is it on reliable news websites? Who posted it first (Grade 7’s reverse image search)?
• Never share a video you think may be fake “just to ask”: sharing spreads it.
• Report it: to the app, and, if someone is cheated or harmed, to Nepal Police’s Cyber Bureau (Unit 8’s Where to get help in Nepal).

In Grade 6 you learned that AI runs in data centers that use a great deal of electricity, and water to cool the computers Grade 6: Use AI only where it helps. Now you can see where the cost comes from:

• Training a large model once takes thousands of computers working for weeks or months.
• Every use costs again. Each answer is made by computers in a data center, word by word. One short answer uses only a little electricity, but chatbots answer billions of prompts.
• Bigger costs more. A long answer costs more than a short one, a bigger model more than a small one, a picture more than a short text answer, and a video far more than a picture. Asking for twenty versions costs twenty times as much.

The electricity comes from power stations, and in many countries much of it still comes from burning coal and gas, which adds to climate change. Nepal’s electricity comes mostly from hydropower, but the data centers that answer your prompts are almost all in other countries.

So the rule from Grade 6 still holds, with numbers behind it: use AI where it helps, and use a simpler tool where one does the job. A calculator for a sum, a dictionary for a spelling, a search for a fact you can find, your own head for your times tables.

Translation apps are generative AI too: they make a new sentence in another language. They learned from millions of sentences that people had already translated, side by side. As Grade 6 said, they do better with languages they have more examples of: there are far more translated sentences for English and French than for English and Nepali, and far fewer again for Maithili, Tamang or Nepal Bhasa. So translation into and out of Nepali makes more mistakes. Some kinds are especially common:

• How polite “you” is. English has one “you”. Nepali has तँ, तिमी, तपाईँ and हजुर, and choosing wrong can be rude. From English, the tool must guess. A letter to a head teacher can come out sounding like a message to a younger brother.
• Who is a man or a woman. Nepali verbs often show it (ऊ गयो, ऊ गई); English often doesn’t (“my friend went”). The tool guesses, often “he”.
• Sayings. A saying (उखान) or idiom translated word by word loses its meaning. “It’s raining cats and dogs” is about heavy rain, not animals; मुखमा राम राम, बगलीमा छुरा is about a two-faced person, not about prayers or knives.
• Names and words with two meanings. A person or place name can be translated as if it were an ordinary word, and a word with two meanings can get the wrong one.
• Things Nepal does its own way: dates in Bikram Sambat, amounts in lakh and crore, school terms like SEE, kinship words like दाइ, भाउजू and माइजू, which English has no single words for.

Putting the unit together, here is when generative AI is the right tool:

• when it helps you think or learn: explaining an idea another way, giving you practice questions, suggesting ideas you then choose from;
• when a mistake is cheap, or easy to check: a first draft you will rewrite, a summary of a page you have, code you will test;
• when it does a job no simpler tool can.

And when it isn’t:

• when a simpler tool does the job exactly: a calculator, a dictionary, a map, a timetable, an ordinary program;
• when a mistake could hurt someone and you can’t check: medicine, money, the law, safety (Grade 6’s three questions: what if it’s wrong, has it seen things like this, can I check it?);
• when it would mean sharing something private, about you or anyone else;
• when the point is to learn to do it yourself: your homework, your test, your diary, your own writing.

AI can be a patient tutor, available at midnight, who never minds explaining again. It can also do your homework for you, and then you have learned nothing, and handed in work that isn’t yours. The difference is not the tool; it is how you use it.

• Your teacher’s and your school’s rules come first. Some tasks allow AI, some allow it for some parts, some don’t allow it at all. If you don’t know, ask before you start.
• Handing in AI’s work as your own is cheating, just like copying a friend’s answers or a website (Grade 6 called it plagiarism). So is using AI in a test where it isn’t allowed.
• Using AI to learn is not cheating: asking it to explain, to quiz you, to give a hint, to find what is wrong with your work, when you then do the work yourself.
• Always say when you used it, and how, unless your teacher says it isn’t needed.

How to say you used AI

Write a short note with your work, an AI-use note, that says:

1. which tool you used, and when;
2. what you asked it (your prompts, or what they were about);
3. what you used from its answer, and what you did yourself;
4. how you checked what it said.

I used a chatbot (its name) on 12 Kartik 2083 to explain how alever works, because I didn't understand page 45 of my book.I asked for three everyday examples. I used one of them (asee-saw) in my answer. The rest of the answer is my own words.I checked the examples against my science book and they agree.

If you quote AI’s exact words, put them in quotation marks and say they came from AI, as you would with any source (Grade 7, Citing what you used). But an AI’s answer is not a source of facts: cite the book or the website where you checked the fact, not the chatbot.

So far this unit has looked at one person using AI: you, with a prompt. But a new technology also changes a whole community, a town, a ward, a village, in ways no single user chose. Generative AI is an emerging technology: one that is new, spreading fast, and whose effects nobody yet fully knows. When nobody knows yet, people argue. Some say it will lift everyone; others say it will take jobs and fill the internet with lies. Both are too simple.

A good way to think it through is a debate: two sides argue for and against a statement, called the motion, in front of judges. A debate is fair when each side gives evidence, not only opinions, and answers the other side’s strongest points, not its weakest.

What counts as evidence

A claim is what you say is true: “AI will take the jobs of typing shops.” Evidence is what shows it: a study, a number from a reliable source, an event that happened, what you saw or were told by the people it happened to. Evidence comes in different strengths:

• Strong: a careful study, with numbers, by people who have nothing to sell; official figures; several reliable reports of the same event.
• Weaker: one person’s story; a company’s claims about its own product; a study of another country, which may not hold in Nepal.
• Not evidence: a feeling, a guess, “everyone says”, a video with many likes, or a chatbot’s answer (it is likely words, not a source).

Weaker evidence still counts, if you say what it is: “One shop owner in our town told me…” is honest, and worth hearing.

Social, cultural, environmental

A technology’s effects on a community are of three kinds, and a fair debate looks at all three:

• Social: people’s work, money, learning, health, safety, and whether they can trust what they see.
• Cultural: language, art, music, stories, customs: the things that make a community itself.
• Environmental: the land, water and air: electricity, water, and the waste of old devices.

Evidence for both sides

Here is evidence on five questions about generative AI in Nepal. Notice that every question has evidence on both sides.

Jobs. Near every district administration office are small shops that type letters and applications, fill in forms and translate documents. Many young Nepalis also earn money online as freelance writers, designers and translators.

• For: a shop owner using AI can draft an English letter or a translation in minutes, check it, and serve more customers; new work appears, such as checking and correcting AI’s Nepali.
• Against: typing, writing and translating are the very tasks generative AI does. A study of the freelance website Upwork found that within months of ChatGPT’s release, freelance writers there got about 2% fewer jobs a month and earned about 5% less; designers lost about as much after image generators arrived. (Grade 7 showed that automation replaces some tasks, changes most jobs and creates others: Machines that take, change and make jobs.)

Farmers. A farmer in a hill village may be a day’s walk from an agriculture technician.

• For: an AI helper can answer at any hour, by voice and in Nepali, about a pest on the tomatoes or how much fertilizer to use, and Nepali companies have started offering this (the box below).
• Against: a chatbot can be confidently wrong, and it knows less about Nepal (this unit); wrong advice about a pesticide can harm crops, people and streams. And many farmers, often older ones, have no smartphone or data, so the help reaches those who already have the most.

Students.

• For: in a careful study in Edo State, Nigeria, in 2024, students who used an AI tutor after school for six weeks, with teachers guiding them, learned much more English than students who didn’t: the researchers put the gain at about one and a half to two years of ordinary learning.
• Against: in a study at a high school in Türkiye in 2023, students who practiced math with a chatbot that simply gave answers did better while they had it, but 17% worse in the test without it than students who never had it: they had let it do the thinking. A version that gave only hints did no harm. How it is used matters more than whether it is used.

The Nepali language and culture.

• For: translation and voice tools let people read and hear things they couldn’t before, and let a Nepali writer reach readers in other languages; AI tools can help keep and teach smaller languages, if people record them.
• Against: AI’s Nepali is weaker than its English, and its Maithili, Tamang or Nepal Bhasa weaker still (this unit, and Grade 7’s bias). A flood of AI-made writing, pictures and songs, cheap to make, can crowd out the work of Nepali writers and artists, and image generators draw Nepal’s people and festivals the way pictures from other countries showed them.

Trust and elections.

• For: AI tools help fact-checkers search and compare pictures quickly, and can put a party’s promises, or the rules for voting, into more of Nepal’s languages. But the evidence on this side is thinner: so far, the harms are better recorded than the help.
• Against: before the House of Representatives election of 5 March 2026, deepfakes and AI-edited pictures of candidates spread on social media, like the fake audio of Swarnim Wagle (above). The Election Commission warned people not to post AI-made content meant to confuse voters, and worked with army and police teams to find and remove false content.

And under every question lies the environment: each answer and each picture uses electricity and water in a data center abroad (this unit’s numbers), while Nepal’s own electricity comes mostly from rivers.