Netflix recommending a show, your inbox catching spam, a car recognising a stop sign: all three are machine learning at work. The name suggests a machine that learns the way we do. It doesn't, and once you see why, it becomes much easier to judge what these systems can and can't do.
Machines don't learn, they fit
When people learn, they take in information, connect it to what they already know and apply it to situations they have never seen. A machine learning model does something much narrower. It is shown a large number of examples and searches for a mathematical function that turns the inputs into the right outputs as often as possible.
Andriy Burkov makes this point in The Hundred-Page Machine Learning Book. A trained model is a function that works well on inputs resembling the data it was trained on. Move far enough away from that data and the output can be completely wrong, with no warning at all. A model trained to read handwritten digits will confidently "read" a handwritten letter as some digit, because digits are the only things it knows exist.
That doesn't make it useless. It means the model is only as good as the data it was fitted to, and it has no idea when it has wandered outside that data.
The three main types of machine learning
1. Supervised learning
Think of a student practising with an answer key. The model gets examples together with the correct answers: thousands of photos labelled "cat" or "not cat", emails labelled "spam" or "not spam". It adjusts itself until its answers match the labels as closely as possible, then applies the same function to new examples. Most of the machine learning inside everyday products is supervised, from spam filters to medical image analysis to speech recognition.
2. Unsupervised learning
Now picture a student handed a pile of books with no instructions, who has to sort them into groups alone. Unsupervised learning gets raw data without labels and looks for structure in it. A typical use is customer segmentation, where a shop groups its customers by buying behaviour without knowing in advance what the groups will be.
3. Reinforcement learning
This is learning by trial and error. The model acts in an environment and receives a reward or a penalty depending on what happens. A game-playing program makes moves, eventually wins or loses, and gradually shifts its strategy toward the moves that lead to wins. Reinforcement learning is behind well-known game-playing systems such as AlphaGo, and it is used in robotics and in fine-tuning large language models.
How a model learns from data: a spam filter
- Collect data. Gather thousands of emails, each labelled spam or not spam.
- Extract features. Turn each email into measurable signals: which words appear, how long it is, how many links it has, who sent it.
- Train. An algorithm such as a decision tree or a neural network searches for the combination of features that best separates spam from everything else.
- Predict. When a new email arrives, the model applies that function and gives its verdict.
- Retrain. Spammers change tactics, so the model is retrained from time to time on fresh, labelled emails.
It is a bit like a student who has done so many practice tests that they recognise the question types. The difference is that the student eventually understands the subject. The model never does. It simply follows the path the optimisation found.
Why the word "learning" misleads
Arthur Samuel, an AI pioneer at IBM, is usually credited with coining the term machine learning in 1959, in his work on a program that played checkers. Burkov argues that it was always more of a catchy label than a literal description. "Artificial intelligence" doesn't mean intelligence in the human sense, and "machine learning" doesn't mean learning in the human sense either. "Automated pattern recognition" or "statistical approximation" would be more accurate, if a lot less exciting.
People and models compared
- People understand context, make leaps of logic and can cope with situations that are completely new to them.
- Models find patterns and optimise their output, but understand nothing beyond what their training data contained.
The consequences are very practical. A driving system trained mostly on sunny roads can struggle in snow it has never seen. Medical imaging models trained on scans from one hospital have repeatedly done worse on scans from another, partly because they had picked up on differences between the scanners rather than on the disease.
What about large language models?
Chatbots built on large language models can feel like the exception, because they write fluently about almost anything. Underneath, it is the same idea at an enormous scale: a model trained to predict the next piece of text, fitted to a large part of the written internet. That scale makes them far more general than a spam filter. They still fail the way every fitted function fails, though. When a question falls outside what their data covered, they don't say "I don't know". They produce something plausible that may be wrong.
Powerful, but still a tool
Machine learning has changed medicine, finance, logistics and entertainment, and it will keep doing so. But it is a tool. A calculator multiplies faster than any person without knowing what the numbers mean, and a model recognises patterns without knowing what they mean either.
Keeping that in mind sets realistic expectations. Good results come from well-chosen data, sensible algorithms and people who check the output, not from hoping a model will understand the problem once it has seen enough data.
So the next time a recommendation seems to read your mind, remember that it hasn't learned anything about you in the human sense. It has found a function that fits people whose behaviour looks like yours.
Inspired by ideas from Andriy Burkov's The Hundred-Page Machine Learning Book.
