What Is Machine Learning? A Plain-English Explanation

Machine learning gets described either as magic or as maths, and neither is much help if you just want to know what it is. This guide is the middle version: what it actually does, how it differs from ordinary software, and what it is genuinely bad at.

What is machine learning, in plain terms?

Machine learning is software that learns from examples rather than from instructions.

Think about how you would write a program to spot spam email. The rule-writing approach means sitting down and listing every rule yourself: flag anything with FREE MONEY in the subject line, flag anything from this list of addresses, and so on. You would be writing new rules forever, because spammers change tactics faster than you can type.

The machine learning approach is different. You collect a few hundred thousand emails that people have already marked as spam or not spam, and you let a program work out for itself what separates the two piles. Nobody writes the rules. The program derives them, and it can derive new ones when you feed it new examples.

That is the whole idea. Everything else is detail about how the deriving works.

How is machine learning different from regular software?

The difference is where the logic comes from. In regular software a person decides the logic and writes it down. In machine learning the logic is discovered from data and stored as a large set of numbers, called a model.

	Traditional software	Machine learning
Who writes the rules	A programmer	The training process, from data
What you give it	Instructions	Examples
How you fix mistakes	Edit the code	Improve the data, or retrain
Can you read the logic	Yes, it is written down	Rarely, it is numbers, not sentences
Behaviour on new cases	Exactly what was specified	An educated guess

That last row matters more than any other. Traditional software either handles a case or it does not. A machine learning model always produces an answer, including for inputs it has no business having an opinion about. It does not know when it is out of its depth.

What are the main types of machine learning?

Most systems fall into one of three categories, defined by what kind of feedback the program gets while it learns.

Type	What it learns from	A everyday example
Supervised learning	Examples with the right answer attached	Spam detection, credit scoring
Unsupervised learning	Data with no answers, it finds structure	Customer segments, anomaly detection
Reinforcement learning	Trial and error with a reward signal	Game-playing systems, robotics

There is a fourth worth knowing because it underpins the AI tools you have probably used: self-supervised learning. The system creates its own questions from ordinary data, for example, hide a word in a sentence and try to predict it. That trick is how large language models are trained, and it is why they could be built from text that nobody had to label by hand.

How does a model actually learn?

The process is repetitive rather than clever. In outline:

1. Collect data. Gather examples relevant to the question you want answered.
2. Split it. Hold some data back so you can test on examples the model has never seen.
3. Make a guess. The model starts with essentially random numbers and predicts badly.
4. Measure the error. A loss function scores how wrong each prediction was.
5. Adjust. Nudge the numbers slightly in the direction that reduces the error.
6. Repeat. Do this millions of times until the error stops improving.
7. Test. Check performance on the held-back data, which is the only honest measure.
8. Monitor. Keep checking after launch, because the world changes.

Step two is the one people skip and regret. If you test a model on the same data it learned from, it will look brilliant and then fail in production, a failure common enough to have a name, overfitting. The model memorised the examples instead of learning the pattern behind them.

Where do you already use machine learning?

Almost certainly several times before lunch:

• The spam folder in your email
• Predicted arrival times in a maps app
• What a streaming service or shop suggests next
• Your bank texting you about an unusual transaction
• Your phone grouping photos by face or location
• Voice-to-text when you dictate a message
• Auto-generated captions on videos

None of these were marketed to you as machine learning. That is the usual pattern: once a technique works reliably, it stops being called AI and starts being called a feature.

What can go wrong?

Four failure modes cause most real-world problems, and none of them are exotic.

Bad or biased training data. A model trained on past decisions will reproduce the patterns in those decisions, including the unfair ones. If a company historically hired one kind of candidate, a model trained on its hiring history will favour that kind of candidate. The model is not prejudiced; it is accurate about a prejudiced record.

Overfitting. Covered above: excellent on the training data, poor on anything new.

Spurious correlations. The model latches onto something that happens to correlate in the training set but means nothing. The classic teaching example is a medical image classifier that learned to detect the ruler that radiologists place next to tumours, rather than the tumours.

Drift. A model trained on last year’s behaviour slowly becomes wrong as behaviour changes. Nothing breaks loudly; accuracy just decays until somebody checks.

Do you need to be good at maths to use it?

To build models from scratch, yes, linear algebra, calculus, and statistics all turn up. To use machine learning in a job, increasingly not. Most practical work now involves choosing an existing model, feeding it your data, and judging whether the output is trustworthy.

The skill that matters most is not mathematical. It is knowing what question you are actually asking, whether your data can answer it, and how you would recognise a wrong answer. People who are good at that are useful with or without the calculus.

How does this relate to the AI tools everyone is talking about?

Machine learning is the broad technique. The chatbots and assistants you have used are a specific product of it: large language models, trained with self-supervised learning on very large amounts of text.

A newer layer sits on top of those models, where the system does not just answer but takes actions across several steps to reach a goal. That is a different thing again, and worth understanding separately, we cover it in our guide to agentic AI.

For more explainers, see our Tech section.