By the end of this lesson you will be able to
- Explain the difference between ordinary software, machine learning and generative AI.
- Describe in plain words how a model learns patterns from examples instead of following written rules.
- Explain how a large language model produces a reply one piece at a time.
- Name three everyday uses of AI and one job it is poorly suited to.
- Say why a fluent answer is not the same as a correct answer.
AI tools now appear in search engines, phones, email, photo apps and school and office software. Many people either fear them or trust them too much. Both reactions come from not knowing what the tool is actually doing. This lesson gives you a simple, honest picture, with no mathematics, so that you can decide when an AI tool will help you and when it will mislead you.
Rules written by people, or patterns found in examples
Most software you have met so far follows rules that a person wrote. A spreadsheet adds numbers because someone wrote exactly how addition works. If you give it the same input twice, you get the same output twice. Programmers call such a set of steps an Algorithm.
Some jobs are too messy for written rules. Try writing exact rules that tell a computer whether a photo shows a cat. Cats come in every colour, pose and light. Instead of writing rules, researchers show the computer thousands of photos already labelled "cat" or "not cat" and let it find the patterns itself. This approach is called machine learning. The examples are the training data, and the result of the learning is a model: a large collection of numbers that captures the patterns.
Artificial intelligence (AI) is the broad name for computer systems that do tasks which normally need human judgement, such as recognising speech, translating or recommending. Today almost all practical AI is built with machine learning. So the three terms nest inside one another, as the picture shows.
Biggest idea at the top
Read this figure as text
Three stacked layers showing artificial intelligence containing machine learning containing generative AI.
Three layers from outside to inside. The outer layer is artificial intelligence: any computer system doing tasks that seem to need human judgement. Inside it is machine learning: systems that learn patterns from examples rather than from hand-written rules. Inside that is generative AI: machine learning systems that create new text, images, sound or code. Large language models, which write text, sit inside generative AI.
How a model learns
Learning here means adjusting numbers. At the start, the model is a blank set of numbers and its guesses are random. It is shown an example, makes a guess, and is told how far off it was. It then nudges its numbers slightly so that the same guess would have been a little better. This happens millions of times. Nobody tells it what a whiskered face or a pointed ear is. It works that out because those features help it get answers right.
Two consequences matter for you. First, a model can only reflect what was in its examples. If the examples were unbalanced, outdated or wrong, so is the model. Second, a trained model is fixed until it is trained again. It does not read the news by itself, and it does not learn from your conversation in the way a friend would.
- Show an exampleFrom the training data
- Model guessesAt first, nearly random
- CompareHow far from the right answer?
- Adjust numbersA tiny improvement
Read this figure as text
A cycle of four steps: show an example, model guesses, compare with the right answer, adjust the numbers.
A repeating cycle with four steps. Step 1: show the model an example from the training data. Step 2: the model makes a guess. Step 3: compare the guess with the correct answer to see how wrong it was. Step 4: adjust the model's internal numbers a little so the guess would be better next time. The cycle repeats millions of times. In the centre is the word training. After training stops, the numbers are fixed and the model is ready to use.
Large language models: predicting the next piece of text
A large language model (LLM) is a model trained on a huge amount of written text. Its basic skill is surprisingly simple: given some text, predict what piece comes next. Imagine the phone keyboard that suggests your next word, but trained on far more writing and able to look at a whole conversation. To answer a question, the model predicts the first piece of the reply, adds it to the text, predicts the next piece, and so on until the reply is complete.
Because the training text contained explanations, stories, code and conversations, the predicted text often reads like a clear explanation, a story or working code. This is called generative AI: it generates something new rather than looking up something stored. The model does not keep a copy of the internet to look things up in. It holds patterns about how language usually continues.
- Step 1: You type a prompt
- Step 2: Model reads itPlus the chat so far
- Step 3: Predicts next piece
- Step 4: Repeats piece by piece
- Step 5: You read the reply
Read this figure as text
A flow from your question to the model predicting pieces of text one after another until a reply is finished.
Five steps. 1: you type a question, called the prompt. 2: the model reads the prompt and the conversation so far. 3: it predicts a likely next small piece of text. 4: that piece is added to the text and the model predicts the next piece, repeating until the reply feels finished. 5: you read the reply. At no point does the model look up a stored fact sheet or check the reply against reality unless the tool is separately connected to search.
Where AI helps and where it does not
AI is a tool, and like any tool it suits some jobs better than others. It is good at tasks where a rough draft is useful and you can judge the result yourself. It is poor where one wrong detail causes harm, or where you cannot check the answer.
| Job | AI suitability | Why |
|---|---|---|
| Summarising a long article you have read | Good | You can compare the summary with the original. |
| Suggesting a clearer wording of your own paragraph | Good | You stay in control and see the change. |
| Translating a short message | Good, with care | Usually accurate, but check names, numbers and tone. |
| Brainstorming ideas for a poster | Good | A weak idea costs nothing; you choose. |
| Working out a medicine dose or a legal deadline | Poor | A wrong detail can harm you. Ask a professional or the official source. |
| Quoting an exact source with page number | Poor | It may invent plausible-looking references. |
| Anything needing today's news or prices | Poor on its own | Without a search connection it only knows its training period. |
You already use AI without a chatbot: spam filters, photo apps that group faces, speech-to-text dictation, map apps that predict traffic, and shop recommendations. These are mostly narrow systems trained for one job. A chatbot is different because one model attempts many jobs, which is why it feels general and why people over-trust it.
Common mistakes
✖ Instead of: Believing a chatbot looks up the answer in a database of facts.
✔ Do this: It predicts text from learned patterns. Unless the tool is connected to a search service, nothing is looked up or checked.
✖ Instead of: Treating AI as one single thing that is either magic or dangerous.
✔ Do this: AI covers many different tools. Ask what a particular tool was built for, and what happens if it is wrong.
✖ Instead of: Assuming a more polished answer is a more accurate one.
✔ Do this: Polish is about language, not facts. Judge accuracy by checking, not by tone.
If something goes wrong
- I cannot tell whether a tool I use is AI.
- Ask whether it improves by learning from examples and whether it makes guesses (recognising a face, predicting a word). If it always follows the same fixed steps, it is ordinary software.
- The terms AI, machine learning and generative AI keep mixing together.
- Picture three nested boxes: AI is the largest, machine learning sits inside it, and generative AI sits inside machine learning. Draw it on paper once.
Key terms
- Artificial intelligence
- Computer systems that do tasks which normally need human thinking, such as recognising speech or writing text.
- Machine learning
- A way of building software that learns patterns from many examples instead of following rules a person wrote.
- Large language model
- An AI system trained on huge amounts of text so it can read and write language.
- Generative AI
- AI that creates new text, images, sound or code instead of only labelling or sorting things.
- Training data
- The collection of examples a machine learning model learns its patterns from.
- Model (AI)
- The result of training: a large set of numbers that captures the patterns found in the training examples.
Practise it for real
Reading is not the same as being able to do it. Complete these on your own computer, then tick only what you have actually done. These ticks are your own honest record: nothing checks them for you.
Spot AI in your day
On paper, list five technologies you or your family use in a normal week (for example phone dictation, spam filter, photo search, map traffic, online shop suggestions). For each, write whether it follows fixed written rules or learns patterns from examples, and what it produces (a label, a prediction or new content). You do not need a computer.
Challenge: explain it in your own words
Without looking at the lesson, write or say an explanation of how a chatbot produces an answer, as if to a friend who has never used one. Include the words training data, pattern, and next piece. Then say one job you would use it for and one you would not, with a reason for each.
Knowledge check
6 questions. You need 80% to pass. Mistakes here are part of learning, and you can retry as often as you like.
Summary
- Ordinary software follows written rules; machine learning finds patterns in examples.
- Training adjusts a model's numbers over millions of guess-and-compare cycles.
- A large language model writes by predicting the next piece of text, again and again.
- AI is useful for drafts and ideas you can judge yourself, and risky where a wrong detail matters.
- A confident, fluent answer is not evidence that it is right.
Checklist: can you do these?
Marking a lesson as read does not prove you can do it. Skills count as demonstrated when you pass the quiz (not yet) and complete the practical tasks (not yet).
Related lessons
References and further reading
Official sources for details that can change between versions.
- Machine learning glossary and guides (Google for Developers)