Part I: Understand AI | Chapter 2

How AI Actually Works

The Goal

By the end of this chapter you will understand, in simple terms, how an AI assistant produces its answers. You will know why it sometimes gives different answers to the same question, why it doesn’t know today’s news, and which mistakes to expect. This understanding will make you a much better user than most people.

Understand: learning from text

An AI assistant is built on a language model. Before you ever talk to it, the model goes through a long training process. It reads an enormous amount of text: books, websites, articles, conversations, code. Far more than any person could read in many lifetimes.

From all this text, the model learns patterns. Which words usually follow which. How a polite letter is structured. How a recipe sounds. How an argument is built. How questions are usually answered.

It doesn’t store the texts like a library. It stores patterns, a bit like how you can speak your language fluently without remembering every sentence you ever heard.

Understand: one word at a time

When you ask a question, the model writes its answer piece by piece. At every step it asks itself: given everything so far, what is the most likely next piece of text?

Try it yourself in your head. Complete this sentence:

At the market, the mama mboga sells tomatoes, onions and …

You probably thought sukuma wiki or spinach. Not laptops. You made a prediction based on patterns you have learned. A language model does the same — billions of times, extremely fast, with patterns from far more text than you have ever seen.

Two things follow from this:

There is often more than one likely next word. The model picks among them with a little bit of chance. That is why the same question can get two different answers. Neither is “the” answer.
The model produces what sounds right. Most of the time, what sounds right is right. But not always.

Understand: from model to assistant

A raw language model only continues text. To become a helpful assistant, it gets a second round of training. People rate its answers and teach it to follow instructions, to be helpful and polite, and to refuse harmful requests.

This training is why the assistant answers your questions instead of just continuing your sentence. It is also part of why it sometimes gives you an answer when the honest answer would be “I don’t know”: it has learned that helpful answers are rewarded.

Understand: what the AI “sees”

Three ideas help you understand what an assistant can and cannot know.

Knowledge cutoff. The model’s training stops at a certain date. It knows nothing about events after that date — unless the tool can search the internet. If you ask about last week’s news without search, it either says it doesn’t know or, worse, guesses.

Context. During a conversation, the assistant reads everything in that chat: your messages, its own answers, any files you added. This is called the context. It only knows about you what is in the context. If you start a new chat, it starts from zero — unless the tool has a memory feature switched on. ChatGPT, Claude and Gemini can remember things about you across chats, also in their free versions. You can see, delete or switch off what they remember in the settings.

Tools. Many assistants can now do more than write. Some can search the web, read documents and photos, listen and speak, or create images. When an assistant searches the web, it can answer questions about current events — but it can still misread or mix up what it finds. The free versions of ChatGPT, Claude and Gemini include web search.

Understand: why Kiswahili sometimes works less well

The models learned from far more English text than Kiswahili text. They can usually understand and write Kiswahili, but the quality is often lower: some phrases sound translated, some terms are unusual, Sheng is mostly unknown.

A practical rule: for important tasks, work in English first, then ask for a Kiswahili version, and have a fluent speaker read it before you use it.

The typical mistakes

Mistake	What it looks like	What you do
Invented facts (hallucinations)	A confident answer with names, numbers or events that don’t exist	Check every fact that matters in a second source
Invented sources	Book titles, links or studies that don’t exist	Open every link. Search for every title.
Outdated information	Old prices, old rules, old office holders	Ask for the date of its information. Check current facts yourself.
Calculation errors	Wrong totals, wrong percentages	Use a calculator for anything that matters
Western default	Examples, prices and laws from the USA or Europe	Always say you are in Kenya and what applies here
Agreeing too easily	“That’s a great idea!” for a weak plan	Ask it explicitly to criticise your idea

You will learn to handle each of these in detail in Chapter 7.

See It in Action

Faith asks an assistant for “five jobs I can do online from Kenya”. She likes the list. The next day she asks exactly the same question on her cousin’s phone — and gets a different list. Only two jobs appear in both.

“Which one is right?” she asks me.

Neither and both. Each answer is one of many likely answers. Faith now does something clever: she asks for fifteen options, then asks the assistant to sort them by how realistic they are for someone with only a smartphone. Different answers have become a source of more ideas.

Kevin is an M-Pesa agent. He asks an assistant to explain the current M-Pesa transaction charges so he can print a table for his customers. The answer looks professional. But when he compares it with the official tariff from Safaricom, several numbers are wrong. The assistant had learned an older table and had no way of knowing it had changed.

Kevin’s rule from now on: official numbers come from the official source. He still uses AI — to design a clear, friendly layout for the official table.

Try It Yourself

Three small experiments. Each takes two minutes.

Same question, twice. Ask any question, then open a new chat and ask it again. Compare the answers.
The cutoff. Ask: “Up to which date does your knowledge go? Can you search the internet in this conversation?”
The trap question. Ask about something that doesn’t exist, for example: “Who won the Golden Cob Award at the Bungoma Maize Festival in 2019?” (There is no such award.) Does the assistant admit it doesn’t know — or does it invent a winner?

Copy & Paste Prompt

Use this whenever the facts in an answer matter.

[Your question]

After your answer, add a section called "How sure am I?".
List:
1. The parts you are confident about.
2. The parts I should check in another source, and where I could check them.
3. Anything that may have changed since your training data.
I am in Kenya. If something depends on the country, say what applies to Kenya.

Improve the Result

If the answer is too general: “Give me a concrete example from Kenya.”
If you suspect it is guessing: “Are you sure? What is this based on?”
If it gave a source: “Give me the exact title and the website where I can find it.” Then check it yourself.

A warning: asking “are you sure?” sometimes makes the assistant change a correct answer, just to please you. Its second answer is not automatically better. The real check is always outside the AI.

Check Your Work

Do I know whether this assistant could search the internet for this answer?
Did the answer contain names, numbers, dates, prices or rules? Did I check them?
Did I tell it that I am in Kenya?
Did I compare at least two answers when the decision is important?

Apply It

Take your goal from Chapter 1. Ask an assistant for advice on it, using the prompt above. Then pick one fact from its answer and check it in a second source: a website, a person who knows, an official office.

Write down: Was the fact correct? Partly? Wrong?

Key Takeaways

A language model learns patterns from huge amounts of text and predicts its answers piece by piece.
Different answers to the same question are normal. Use them as extra ideas.
The model has a knowledge cutoff. Without web search it doesn’t know current facts.
It only knows about you what is in the conversation — and in its memory, if that is switched on.
Expect typical mistakes: invented facts, outdated information, Western defaults, too much agreement. Official facts come from official sources.