Machine learning is the part of artificial intelligence in which software learns from examples instead of a person programming every rule by hand. The result, the model, recognises patterns in data and uses them to make predictions, such as whether an email is spam, or to generate new content such as text and images. ChatGPT and Claude are built on it too: behind them are large neural networks with many layers, and learning with such networks is called deep learning.
I work with Claude every day. My fourteen AI employees run on Claude Team, some also on open models or through the API. The terms in this article follow Google's introductory course on machine learning and Anthropic's glossary, both retrieved on 27 October 2026. All the basics about AI are collected on my AI Basics page.
How it differs from a classic program
A classic program works with rules that a person writes down: if the subject line contains "prize", move the email to spam. That works until senders change their subject lines. Then someone has to write new rules.
Machine learning turns this around. You give the software a great many examples, here thousands of emails labelled "spam" or "not spam". The software looks for the patterns that separate the two groups on its own. Google describes machine learning exactly like this: as the process of training a piece of software, called a model, to make useful predictions or generate content from data.
Google gives a vivid example. To predict rainfall, you could rebuild the physics of the atmosphere in huge equations, which is incredibly difficult. Or you give a model enormous amounts of weather data until it has learned the relationship between weather patterns and rainfall itself. Then it gets the current data and predicts how much rain will fall.
How the learning works
Three words are enough to follow almost any explanation:
- Features are the inputs the model works with. For an email, for example, the sender, the subject and the words in the text.
- Labels are the correct answers in the training examples, here "spam" or "not spam".
- Training is the process in which the model adjusts its internal settings until it fits the examples. According to Google, it uses each example anywhere from a few times to billions of times.
One trap has its own name: overfitting. The model then fits its training examples so closely that it gets new data wrong. Like an apprentice who has memorised old exam questions and fails at a new one.
The four types of machine learning
Google divides machine learning systems into four groups. Some of the examples in the right-hand column are my own made-up ones; spam, the robots and AlphaGo are Google's own examples:
Within supervised learning, Google distinguishes two tasks. Regression predicts a number, such as a house price or a travel time. Classification predicts a category, such as spam or not, rain or no rain. Google compares supervised learning to a student preparing for an exam by working through old exams that include the answers. What generative AI can do in detail is covered in my article on generative AI.
Neural networks and deep learning
A neural network is a particular kind of model. It consists of many small computing units arranged in layers that are passed through one after the other. The layers between input and output are called hidden layers. Google defines a neural network as a model with at least one hidden layer. If it has more than one, it is called a deep neural network, and learning with it is called deep learning.
This gives a family tree in which each level is part of the one above:
- Artificial intelligence is the umbrella term. Google describes AI as a program or model that solves sophisticated tasks, such as translating text or identifying diseases in radiology images.
- Machine learning is, according to Google, a sub-field of AI: the part that learns from data.
- Deep learning is machine learning with deep neural networks.
- Large language models, LLMs for short, are language models with a very high number of parameters, meaning learned settings. They are usually based on a neural network architecture called the Transformer, which consists of many layers. Behind ChatGPT, Claude and Gemini there is such a model in each case.
Anthropic describes the last step like this: in pretraining, the language model learns to predict the next word given the text before it. It is then trained further with additional data and with feedback from people to become a helpful assistant. How that works in detail is in How Do LLMs Work? A Plain-English Guide.
Where you meet machine learning every day
Google names translation apps, weather forecasts, estimated travel times, song recommendations and sentence auto-completion. At the office you know it from the spam filter and from every AI chat. Most of the time you do not even notice that a trained model is working in the background.
What to watch out for
- A model is only as good as its examples. If it learns from one-sided or faulty data, it takes on those errors.
- A prediction is not a certainty. A model says what is likely according to its data. With language models, that can be an answer that sounds convincing but is wrong. More on that in AI Hallucinations: Why AI Makes Things Up.
- You do not need to train your own model. My assessment for small businesses: the question is rarely how to build a model, but which process a ready-made tool can help with. A model of your own only pays off when you have a lot of your own data and a clearly measurable task.
Frequently asked questions
What is the difference between AI and machine learning?
AI is the umbrella term. Machine learning is part of it, namely the part where software learns from data instead of only following fixed rules. According to Google, some organisations now use the two terms interchangeably, but strictly speaking machine learning is a sub-field. The AI tools everyone is talking about, such as ChatGPT or Claude, are based on machine learning.
What is the difference between machine learning and deep learning?
Deep learning is a form of machine learning. It uses neural networks with more than one hidden layer. Large language models such as the ones behind ChatGPT and Claude are deep networks of this kind.
Is ChatGPT machine learning?
Yes. ChatGPT is based on a large language model, a deep neural network that has learned from huge amounts of text. The same applies to Claude and Gemini.
Do I need to know how to code to use machine learning?
To use it, no. Every AI chat, spam filter and translation app uses trained models without you writing a line of code. Only if you want to train models yourself do you need programming and statistics skills.
What are neural networks, explained simply?
A neural network is a model made of many small computing units arranged in layers one after the other. Each layer processes what the previous one outputs. During training, the connections between the units are adjusted until the network fits the examples.
Where to go from here
This week, notice where you meet trained models: in your inbox, in the translation app, in the AI chat. Each time, ask yourself which examples this model probably learned from. The step before this, what generative AI is, is covered in Generative AI Explained, with Examples. Next comes the question of what an AI model is and how it differs from the app you use. You will find every technical term from this article briefly explained in AI Terminology: A Plain-English Glossary. How others use AI in their business is what we discuss in my community Claude Practitioners.
