Lesson 1.2: What Is Generative AI?

Courses › AI Foundations for Everyday Work › Module 1: Understanding AI › Lesson 1.2

Module 1 · Lesson 2 of 3Reading time: about 9 minutesPractice: about 15 minutesNeeds any AI assistant for the exercise

What you will be able to do

Explain in plain words how a generative AI assistant produces text and images, and complete one real task: give an instruction, check the result, and improve it.

What “generative” means

Most AI you met in Lesson 1.1 sorts or spots: junk or not junk, face or no face. Generative AI creates. You describe what you want and it produces new text, images, audio, video or computer code. It can also transform something you give it: summarise a report, translate a message, rewrite a paragraph in a friendlier tone, or restyle a photo.

1. You giveAn instruction and the facts that matter: a task, your notes, who it is for.
2. AI producesA draft built from the patterns it learned, shaped by what you gave it.
3. You checkFacts, names, numbers, tone and promises. Then you improve it or use it.
Diagram: the working loop for every task in this course. Examples: a message draft, a summary of a long document, or the idea for an image. Conceptual illustration.

How it produces an answer

A generative model is trained on a very large collection of examples: text written by people, or pictures paired with descriptions. From these it learns patterns: how sentences usually flow, what a customer reply usually contains, what a “shop at sunset” picture tends to look like.

When you ask for something, a text assistant builds its reply one small piece at a time, each time choosing a piece that fits your instruction and the patterns it has learned. An image tool works differently inside, but the idea is the same: it uses patterns learned from examples to make something new that matches your description.

Think of someone who has read an enormous library and learned how writing usually goes. They can draft quickly in almost any style. But they are writing from patterns they learned, and they have never met your customers. Some assistants answer only from what they learned in training. Others can also search the web or use connected sources, and a good one shows where it looked. Either way, a source is something for you to open and check, not proof.

Three things follow from this:

  • It sounds confident whether or not it is right. A fluent sentence is not evidence.
  • Ask twice and the wording may differ. That is normal. It is not a mistake you caused.
  • It only knows your situation if you say so. Its training also stops at some date, so recent events and current prices are not reliable unless the tool can look things up and shows you where it looked.

Generated is not the same as verified

Treat every AI answer as a draft. A generated sentence can read exactly like a fact. Names, dates, figures, quotations, sources and promises must be checked against something you trust before anyone else sees them.

A real task, step by step

Training example The people and workshop below are invented. The conversation is real: the instruction in step 2 was typed into an AI assistant (Claude, from Anthropic) on 12 November 2026, and the replies are shown exactly as it wrote them. Your assistant will word things differently. That is expected.

This lesson shows the conversation as text so you can read it closely and copy the prompt. How to open and use an assistant’s screen is the subject of Module 2.

The starting problem

Sam runs a small furniture repair workshop. A customer, Ms Daniels, was quoted by phone for fixing a loose joint on her bookshelf. Inspecting it, Sam found a split back panel that also needs replacing. She has to be told, and her approval is needed before the extra work. Sam has rough notes but does not want to spend time wording a polite message.

The instruction given to the AI

Sam pastes in the notes, says what to write and for whom, and sets a limit on what the AI may add.

Sam’s instruction (the prompt)
I run a small furniture repair workshop. A customer, Ms Daniels, phoned and I quoted her over the phone to fix a loose joint on her oak bookshelf. When I inspected it, I found the back panel is also split and needs replacing. Here are my notes:

- Phone quote was only for the loose joint.
- The back panel needs replacing too: about 2 extra hours of work plus one new panel.
- I have not priced the extra work yet. I will send a revised quote tomorrow.
- I need her OK before I carry on with the panel.
- Her bookshelf stays in my workshop meanwhile. No work on the panel until she approves.

Write a short, friendly WhatsApp message to Ms Daniels explaining this. Do not add any prices, dates or promises that are not in my notes.

Notice three things: the facts come from Sam, not from the AI; the message type and tone are named; and there is an instruction about what not to add.

The first result

The assistant’s first reply
Hi Ms Daniels, it's [your name] from [workshop name]. I've had a proper look at your oak bookshelf. The loose joint is fine to fix as we discussed on the phone, but I also found that the back panel is split and needs replacing. That would mean about 2 extra hours of work plus a new panel, which wasn't part of the phone quote.

I haven't priced the extra work yet. I'll send you a revised quote tomorrow.

I won't do anything on the back panel until you've had a look at the new quote and given me the OK. Your bookshelf is safe in my workshop in the meantime.

Any questions, just let me know. Thanks!

What to check

Sam reads the draft against the notes, line by line. The draft is good, and the check still finds four things to fix:

  • Blanks to fill. It left [your name] and [workshop name]. It cannot know them. Sam types them in.
  • Something not in the notes. “safe in my workshop” is a promise about the bookshelf that Sam never wrote. “Staying in my workshop” is what the notes say.
  • Unclear wording. “The loose joint is fine to fix” could be read as “it is fixed” or “it might not be fixable”. The notes only say it is covered by the phone quote.
  • No clear question. Sam needs a yes or no. The draft ends with “Any questions”, so the customer might not know what to answer.

The things Sam asked it not to invent (prices and dates) are absent. Sam still checked, because asking is not a guarantee.

How the result was improved

Sam did not start again. Sam asked for a revision, saying exactly what to change:

Sam’s follow-up
Thanks. Please revise the WhatsApp message. Make it shorter. Replace "safe" with plain wording that only repeats my notes. Make it clear the loose joint is covered by the phone quote and the panel is not. End with one clear yes-or-no question she can answer once she has seen the revised quote. Reply with the revised message only, still no prices, dates or promises that are not in my notes.
The assistant’s revised reply
Hi Ms Daniels, it's [your name] from [workshop name]. I've inspected your oak bookshelf. The loose joint is covered by the phone quote. The split back panel is not, because it needs replacing, which means about 2 extra hours of work plus a new panel.

I haven't priced that yet, and I'll send you a revised quote tomorrow. Your bookshelf is staying in my workshop, and I won't do any work on the panel until you approve.

Once you've seen the revised quote, can I go ahead with the back panel, yes or no?

Sam checks again: the “safe” promise is gone, the covered and not-covered parts are clear, and the message ends with one question. Sam fills in the two blanks and sends it. Sam is the one who sends it, so Sam is responsible for it.

What this example shows. The AI saved the work of wording the message. It did not replace Sam’s facts, Sam’s checking, or Sam’s decision to send. A second request that names the exact changes usually works better than starting over.

Common mistakes

  • Giving no facts. “Write a message to a customer about a problem” invites invented details.
  • Skipping the limits. Say what not to add: prices, dates, promises, claims.
  • Sending the first draft. Always read it against your notes.
  • Asking for everything at once. One task, then one revision, works better than a long list.
  • Pasting private details. Use “a customer” or “[customer]” instead of real names, numbers and addresses. Lesson 1.3 explains why.

Try it yourself: your own checked draft

Pick a small, real writing task from your own life that needs no private information: a message to a customer, a note to a classmate, a reminder to a colleague, or a short description of your service. Use any AI assistant you can reach: a chat tool such as ChatGPT, Claude, Gemini or Microsoft Copilot. To paste, copy the text, click into the assistant’s message box, paste it, then press Send. If you do not have one yet, do steps 1 and 2 on paper now and run it after Module 2. Reading the demonstration above is not the same as doing the task yourself, so come back and complete the hands-on steps.

  1. Write your facts. Three to five bullet points of what is true. Leave out names, phone numbers and addresses.
  2. Write the instruction. Copy the template below and fill in the brackets.
    Template: draft from my notes
    I am [your role, for example: a shop owner / a student / a freelancer]. Here are my notes:
    
    - [fact 1]
    - [fact 2]
    - [fact 3]
    
    Write a short, [friendly / formal] [message / note / description] for [who it is for]. Do not add any prices, dates, names or promises that are not in my notes.
  3. Read the reply slowly. Tick each fact against your notes. Highlight anything that is not in your notes, anything unclear, and any blanks.
  4. Ask for one improvement. Name exactly what to change, as Sam did.
  5. Decide. Edit it yourself and, if it is something you would send, use it. If it is not good enough, say why in one line.
  6. Record it in the worksheet, Part B (task, instruction, what you checked, what you changed).

You are done when you have two versions, can point to at least one thing you checked, and can say what you changed and why. If the assistant gave a perfect first draft, check once more: that is when mistakes slip through.

Key takeaway

Generative AI creates drafts from patterns it learned, so it is fast and fluent but not a source of verified facts. You supply the facts, set the limits, check the result, and ask for specific improvements.