Learning from examples instead of rules
Machine learning is a way of getting computers to perform a task without writing explicit instructions for it. Instead of a programmer specifying every rule, you show the system many examples and let it work out the pattern. Want to detect spam? Rather than listing every suspicious phrase by hand, you feed a model thousands of emails already labelled “spam” or “not spam,” and it learns which features predict which label. The model is essentially a function with millions of adjustable numbers (parameters); learning is the process of tuning those numbers until the function’s predictions match the examples.
The difference from traditional programming
The cleanest way to grasp machine learning is by contrast. Traditional software is rules in, answers out: a human encodes the logic, the computer executes it. Machine learning flips this to data in, rules out: you supply example input-output pairs, and the algorithm produces the rules (the trained model) that best reproduce them. This is powerful precisely where rules are too numerous or fuzzy to write down — recognising a cat in a photo, transcribing speech, or predicting which customer will churn. Nobody can articulate the exact rule for “this pixel arrangement is a cat,” but a model can learn it from enough labelled images.
How a model actually learns
Training is an optimisation loop. The model makes a prediction, a loss function measures how wrong it was, and an algorithm (gradient descent) nudges every parameter slightly in the direction that reduces the error. Repeat this across the dataset many times and the parameters converge toward values that make good predictions. The critical, often-missed point: the model only learns patterns that exist in its training data. Coverage and quality of that data dominate everything — a model trained on biased or narrow data will faithfully reproduce those flaws, which is why “garbage in, garbage out” is the iron law of the field.
The three main types
Machine learning splits into three families. Supervised learning uses labelled data — each example comes with the correct answer — and powers most practical systems: fraud detection, price prediction, image classification, and the fine-tuning of language models. Unsupervised learning works with unlabelled data and finds hidden structure, such as grouping customers into segments or spotting anomalies. Reinforcement learning has an agent learn by acting in an environment and receiving rewards or penalties, the approach behind game-playing systems and part of how chatbots are aligned to human preferences.
Why “accuracy” is not the whole story
A model can score high accuracy and still be wrong in dangerous ways. It might have latched onto a spurious correlation — predicting “dog” because every dog photo in training had grass, not because it recognised the animal. It captures correlation, not understanding or cause, so it can fail abruptly on inputs unlike its training set. The practical takeaway for anyone using ML: a good score is necessary but not sufficient. Always ask what data the model learned from, where it is likely to fail, and whether it learned the pattern you actually care about.