What Is Machine Learning? How Computers Learn From Data

What Is Machine Learning?

Machine learning is the process of training software to make predictions by learning from data. Google’s introduction to the subject defines it as “the process of training a piece of software, called a model, to make useful predictions”, and notes that models can now generate content too, such as text, images and audio.¹

The key idea is that nobody writes the rules directly. To build a spam filter the traditional way, a programmer would list warning signs: certain words, suspicious links, odd senders. A machine learning spam filter is instead shown thousands of emails already marked “spam” or “not spam”, and it works out for itself which patterns separate the two.

Machine learning is a branch of artificial intelligence. Most of the AI people use today, from photo search to large language models, is built with it.

Machine Learning vs. Traditional Programming

Traditional programs follow rules people write; machine learning programs learn rules from examples. Google illustrates the difference with weather. A traditional approach to predicting rainfall would model the physics of the atmosphere, which is extremely hard. A machine learning approach gives a model large amounts of past weather data and lets it learn how conditions relate to rain, then uses today’s conditions to predict.¹

Neither is better in general. If the rules are known and must be exact, such as calculating income tax, ordinary programming is simpler and more reliable.

A Short History: Teaching a Computer to Play Checkers

The idea is older than most people think. In 1959, IBM researcher Arthur Samuel published a paper in the IBM Journal showing that a computer could be programmed to “learn to play a better game of checkers” than the person who wrote the program.³ The program improved by playing, adjusting the weights it gave to different features of a position, and reached that level after about 8 to 10 hours of playing time.³

Samuel’s work is widely credited with popularising the term “machine learning”. For decades the field grew slowly, limited by computing power and data. The explosion since the 2010s came from three things arriving together: far more digital data, much faster chips, and better methods, especially deep neural networks.

How Machine Learning Works, Step by Step

Most machine learning projects follow the same basic cycle.

1. Collect examples. Gather data that represents the problem, such as past emails or house sales. In many projects each example has a known answer, called a label.
2. Split the data. Keep some examples back for testing. Google’s course shows an illustrative split of 70% for training, 15% for validation and 15% for final testing, and stresses shuffling the data first so each part is representative.²
3. Train. The model makes predictions on the training data, compares them with the right answers, and adjusts its internal settings to reduce its mistakes. It repeats this many times.
4. Test. Check the model on data it has never seen. This is the honest measure of how it will perform in the real world.
5. Use and monitor. Put the model to work, keep checking its accuracy, and retrain it as the world changes.

The Main Types of Machine Learning

Google groups machine learning systems into four broad types.¹

Supervised learning is the most common in business. It splits into two kinds: regression, which predicts a number such as a price, and classification, which predicts a category such as “fraud” or “not fraud”.¹

A Simple Example: Predicting House Prices

Imagine teaching a model to estimate house prices. You give it past sales, each with features such as floor area, number of bedrooms and location, and the price each house actually sold for.

At first the model’s guesses are poor. With each pass through the data it nudges its settings: perhaps each extra square foot adds a certain amount, a better neighbourhood adds more, an older roof subtracts a little. After training, you show it houses it has never seen and compare its estimates with their real prices. If it is usually close, it is ready to use; if not, you add better data or try a different kind of model.

Notice what the model needs to do well: enough past sales, prices that reflect today’s market, and features that actually matter to buyers. A model trained only on sales from one city will struggle in another, however clever its method.

No one told the model what a square foot is worth. It worked that out from the examples. That is both the power of machine learning and its risk: it learns whatever patterns are in the data, including ones you didn’t intend. This example is an illustration of the general method, not a real model.

Where You Already Use Machine Learning

Machine learning sits behind many everyday services, usually unnoticed.

• Email: spam filtering and suggested replies.
• Banking: flagging unusual card payments that may be fraud.
• Shopping and streaming: recommendations based on what similar people chose.
• Phones: face unlock, voice assistants and photo search.
• Maps: traffic predictions and arrival times.
• Search and chat: ranking results and generating answers; see how Google’s AI Overviews use it.

Where Machine Learning Goes Wrong

A model is only as good as its data and its testing. The most common problems:

Overfitting

Google defines overfitting as creating a model that “matches (memorizes) the training set so closely” that it fails on new data.² It is like a student who memorises past exam answers without understanding the subject: perfect on old papers, lost on a new question. Testing on held-back data is how it is caught.

Biased data

If past data reflects unfair decisions, a model trained on it can repeat them. A hiring model trained on a company’s past hires may learn to favour the kinds of people it hired before. The US National Institute of Standards and Technology’s AI Risk Management Framework treats harmful bias as one of the risks organisations should identify and manage.⁴

A changing world

Models learn from the past. When customer behaviour, prices or fraud tactics change, accuracy can quietly drop, which is why models need monitoring and retraining. This is sometimes called drift. A demand forecast trained before a big price change, for example, may keep predicting old buying patterns until it is retrained on newer data.

Too little or poor-quality data

Missing values, duplicates, wrong labels and data that covers only some situations all weaken a model. Teams often spend more time collecting, cleaning and checking data than building the model itself, and that effort usually matters more to the result than the choice of algorithm.

Confident mistakes

Generative models can produce fluent answers that are simply wrong. Our explainer on why AI makes things up covers this in detail.

Machine Learning in Business

Most business uses of machine learning are about predicting something a little better than a rule of thumb can. Common examples:

The value comes from acting on the prediction: contacting customers likely to cancel, ordering the right stock, blocking a suspicious payment. A model that is accurate but never changes a decision is just an expensive report. Each of these also needs a person deciding what a wrong prediction costs, because a missed fraud and a blocked genuine customer are both mistakes with different prices.

Is Machine Learning Always the Right Tool?

Often it isn’t, and it is worth checking before starting a project. Machine learning makes sense when there is plenty of relevant data, the pattern is too complex to write as rules, and some errors are acceptable. It is a poor fit when:

• The rules are already known and must be exact, such as tax or legal calculations.
• There is little data, or the data doesn’t reflect the situations the model will face.
• Every decision must be fully explained, and the model can’t explain how it decided.
• A wrong answer would be dangerous and there is no person checking the output.

A useful habit is to build the simplest possible baseline first, such as “predict the same as last month”. If a machine learning model can’t clearly beat it, the simpler approach wins.

Machine Learning, Deep Learning and AI: How They Relate

The terms nest inside one another. Artificial intelligence is the broad goal of computers doing tasks that seem to need intelligence. Machine learning is the main way that goal is achieved today. Deep learning is a kind of machine learning that uses neural networks with many layers, and it drives most recent breakthroughs in images, speech and language.

Large language models are deep learning models trained on huge amounts of text, and AI agents build on them by adding tools and planning. Each step adds capability, and new ways to go wrong.

Machine Learning and Your Privacy

Because models learn from data, the data they learn from matters to you. Services may use your clicks, purchases, messages or uploads to train or improve their models, depending on their terms and the law where you live. Check privacy settings in apps you use often, think twice before pasting personal or confidential information into AI tools, and read how a service says it uses your data if that matters to you.

How to Get Started With Machine Learning

You don’t need a PhD to understand or use machine learning. A sensible path:

1. Learn the ideas first. Free introductions such as Google’s machine learning courses explain the concepts without heavy maths.¹
2. Try a no-code tool. Spreadsheet add-ons and cloud tools let you train a simple model on your own data, with no code at all.
3. Learn some Python. It is the most common language for machine learning, with free libraries for nearly every task.
4. Start with a real, small problem. For a business, that might be predicting which customers are likely to cancel.
5. Measure honestly. Always test on data the model hasn’t seen, and compare with a simple rule of thumb.

Key Terms Explained

These are the words you will meet most often in machine learning.

The Bottom Line

Machine learning lets computers learn from examples instead of rules. It works best on patterns too complex to write down, such as images, language and fraud, and it depends on good data and honest testing. Its weaknesses, overfitting, bias and confident errors, are manageable when people check the results. To see where it fits, read about AI, large language models and agentic AI, or browse more tech explainers.

Keep Reading

Tech & AI

What Is Retrieval-Augmented Generation (RAG)? How It Works

Retrieval-augmented generation (RAG) lets AI look up trusted documents before it answers. Learn how RAG works,…

Tech & AI

What Is a Password Manager? How It Works and Is It Safe?

A password manager creates, stores and fills strong, unique passwords for every account. Learn how it…

Tech & AI

ChatGPT vs Claude vs Gemini: How the Big AI Assistants Compare

ChatGPT vs Claude vs Gemini compared on plans, prices, features, strengths and privacy in late 2026,…