Programme / 5.9 AI Literacy for Teachers

Switch course ▾

5.1 Understanding How Students Learn 5.2 Lesson Planning Fundamentals 5.3 Classroom Management 5.4 Student Engagement 5.5 Assessment Fundamentals 5.6 Differentiated Instruction 5.7 Communication with Students & Parents 5.8 Digital Teaching Fundamentals 5.9 AI Literacy for Teachers 5.10 Reflective Teaching Practice 6.1 Multilingual Classrooms: Uzbek, Russian & English Instruction 6.2 Teaching Large Classes Effectively 6.3 Early Literacy & Numeracy Foundations 6.4 Classroom Technology Basics 6.5 Working with Parents in a Multicultural Community 6.6 Wellbeing & Stress Management for New Teachers Programme (64) →

Session 1. How Generative AI Works

Objective. Explain in plain terms how a generative AI tool produces text, and why the same prompt can give different answers.

Why this matters

When a colleague says 'the AI told me', it helps to know what actually happened inside the tool. You do not need technical knowledge, only a clear picture: the tool continues text in a likely way, using patterns from its training. With that picture you can decide quickly when a draft is safe to use, when to ask again, and when to put the tool away. This session gives you that picture and shows how small changes in your prompt change what you get back.

Key ideas

Prediction, not understanding

A text AI is trained on a very large amount of writing and learns which words tend to follow which.
When you write a prompt, it predicts the most likely next piece of text, then the next, one step at a time.
Fluent does not mean true. The tool is built to sound right, not to check what is right.

Tokens and the context window

The tool reads text in small pieces called tokens — parts of words, not whole sentences.
It can only 'see' a limited amount of your conversation at once. In a long chat, early details can be lost.
What you put in the prompt shapes everything that comes out: clear context in, better draft out.

Why answers vary

The tool does not always pick the single most likely word; it samples from likely options, so each run differs.
Variability is useful for brainstorming and risky for facts: the same question can return a right and a wrong answer.
Treat each output as one draft among many possible drafts, never as 'the' answer.

Where it fails: hallucination and bias

A 'hallucination' is a confident statement that is false — invented facts, sources, quotations or numbers.
The tool reflects patterns in its training texts, so it can repeat stereotypes and is weaker on smaller languages and local contexts.
It has no knowledge of your students, your curriculum or today's news unless you tell it.

How a text tool writes

1Read the promptSplit into tokens

2Score optionsWhich token is likely next?

3Pick oneSampled, not always the top

4Add and repeatThe new token joins the context

The loop repeats for every new piece of text, so an early wrong choice can carry forward.

Training and tuning: where the confidence comes from

Building a text AI usually has two stages. First comes training on a huge collection of writing, where the model learns patterns of language and of how topics are usually discussed. Then comes tuning, where people rate answers and the model is adjusted to be helpful, polite and well organised. This is why the tool sounds calm and certain even when it is wrong: a confident tone was rewarded, but nobody can check every fact.

Two practical consequences follow. The model works from patterns learned up to a certain date, so recent events, new regulations or this year's textbook changes may be missing. And it is not a database: it does not look up a fact and report it, it generates text that resembles how such facts are usually written.

Your prompt is the steering wheel

Because the tool continues the text you give it, the more useful material you supply, the less it has to guess. Four things help most: who the audience is (Grade 6, Uzbek-medium), what the task is, what format you want (five questions, a table), and which source it should use. If you paste a paragraph from the textbook and write 'use only this text', the tool works from the paragraph instead of from memory, and checking becomes much easier.

You can also ask the tool to show its steps or to say what it is unsure about. These requests do not make it truthful, but they give you something to inspect. If the first answer is weak, change the prompt and run it again.

Examples from the classroom

Three answers about photosynthesis

Scene. Dilnoza teaches Grade 5 in Samarkand. She asks an AI tool three times for a four-sentence explanation of photosynthesis. All three answers read smoothly. Two say that plants take in carbon dioxide and release oxygen; the third says plants 'breathe in oxygen to make food'. The third sounds just as confident as the others. Dilnoza compares all three with the textbook page before choosing one for her class.

Insight. Each run is one sample, and a fluent sentence can still be false. Where runs disagree, that is the place to check.

A source that does not exist

Scene. Aziz teaches Grade 9 history in a Bukhara school. He asks the tool for a quotation from a historical figure, with the book title and page number. The answer arrives with an author, a title and a page, all neatly formatted. Aziz searches the school library catalogue and the internet and finds no such book. He decides to use only quotations he has seen in a real book, and tells his students why.

Insight. A hallucination often looks like a real reference. Ask for sources only so you can open them, never to trust them.

The forgotten instruction

Scene. Malika teaches Grade 7 geography in a Russian-medium school in Nukus. At the start of a long chat she asked for short sentences and simple words for a lesson on the Aral Sea. After many exchanges, the answers become long and full of difficult terms. She starts a new chat, pastes the textbook paragraph, repeats her instruction in one line, and the level is right again.

Insight. The tool sees only a limited window of the conversation. Restate key instructions, and start fresh for a new task.

Common pitfalls

Believing that a calm, confident tone means the answer is correct.
Using one run as proof; if two runs disagree, that is a signal to check, not to pick the nicer one.
Asking for sources and trusting them without opening them.
Assuming the tool knows your curriculum, your students or recent changes.

Hands-on activity · ~35 min

Same Prompt, Three Answers

Read the prompt: 'Explain photosynthesis to a Grade 5 class in four sentences.' Three different AI answers (A, B, C) are in your Activity Kit.
Alone (5 min): list what is the same in all three and what differs — length, tone, examples, accuracy.
In pairs (10 min): find any statement you would not give to a class without checking. Mark it.
As a group (10 min): explain in your own words why one prompt produced three answers. Use the words 'predict' and 'sample'.

Your facilitator will guide you through this activity in the session.

Questions to discuss

Which of the three answers would you trust most, and what would you still check before using it?
What does variability change about how you should use these tools for facts versus for ideas?
How would you explain 'it predicts text' to a colleague in one sentence?

Reflect

Where have you already met AI-written text without noticing?
What is one thing you assumed AI 'knows' that you now see it only predicts?
Which of your tasks needs one reliable answer, and which benefits from many options?

Try it tomorrow

Run the same prompt twice before a lesson and underline every place where the two answers disagree; check those first.
Paste one textbook paragraph with 'use only this text' and compare the result with an answer to the same question without it.
Tell students in one sentence, 'The tool predicts likely words; I check the facts', and ask each to verify one claim.