2. How AI learns from data

Think of how a small child learns to recognise a cat. You do not hand the child a book of rules that says 'four legs, fur, whiskers'. Instead the child sees many cats — in the house, on the street, in pictures — and slowly their brain builds a pattern. After seeing enough cats, the child can spot one they have never seen before.

AI learns in almost the same way. Instead of a human brain, there is a program. Instead of a child's eyes, there is data. The program studies a huge collection of examples and finds the hidden patterns inside them. This process is called training.

Data is simply stored information — numbers, words, photos, clicks, or anything that can be recorded. Here are real examples from everyday life in Ethiopia:

• TeleBirr records millions of transactions: who sent money, how much, when, and to whom. That is data.

• A hospital in Addis Ababa stores patient records: age, symptoms, test results, medicines given. That is data.

• A small shop in the Merkato keeps a notebook of what sold each day. Even that paper list is data.

• Every time you search something or tap an app, you create data.

AI needs a lot of this kind of information — usually thousands or millions of examples — before it can learn reliably.

Not all data is the same. There are two main types:

1. Labelled data — each example comes with the right answer attached. For example: a photo of a dog with the tag 'dog', or a customer message tagged 'complaint'. The AI uses these labels as a teacher.

2. Unlabelled data — examples with no tag. The AI has to find patterns on its own. This is harder but sometimes the only option when labelling is too expensive.

Labelling is often done by real people. Imagine Almaz sitting at a computer, looking at thousands of photos and typing 'cat' or 'dog' under each one. That human work is what makes AI possible.

Once training is done, the AI has learned to recognise patterns. A pattern is a connection between things that repeats. For example:

• Every time Abebe buys flour and oil together, he also buys sugar. A shop AI can learn this pattern and suggest sugar at checkout.

• Whenever a message contains the words 'free prize click now', it turns out to be spam. An email AI can learn this pattern and block such messages.

• If a customer service chat ends with a sad face emoji, the next message is almost always a complaint. A business AI can learn this and alert staff.

When the AI meets a new situation, it uses the patterns it learned to make a prediction. It does not truly understand — it is matching what it sees to what it has seen before.

There is a famous saying in computing: 'garbage in, garbage out.' It means that if you feed an AI poor or unfair data, you will get poor or unfair results.

Here are three common data problems:

1. Incomplete data — the data is missing important groups. If a medical AI is trained only on records from male patients, it may miss symptoms that are different in women.

2. Outdated data — the world changes but the data does not. An AI trained on prices from five years ago will give wrong advice today when everything costs more.

3. Biased data — the data reflects old prejudices. If historical hiring records show mostly men in management roles, an AI trained on that data may unfairly reject women's applications.

Understanding these problems does not require coding. It just requires asking the question: 'Where did this data come from, and does it represent everyone fairly?'