What machine learning actually does

Machine learning is discussed constantly and explained rarely, which leaves most people with a vague sense that computers have started thinking. A clearer picture is available without any mathematics, and it makes both the capabilities and the failures considerably easier to predict.

The basic move

Traditional programming works by instruction. A developer writes rules, the computer follows them. To detect spam this way, you would write out the conditions — certain words, certain senders — and the program would apply them.

Machine learning inverts this. Instead of rules, you supply examples: a large collection of messages labelled spam or not spam. A learning algorithm searches for statistical patterns distinguishing the two groups, and the result is a model that can classify messages it has never seen.

Nobody wrote the rules. Nobody, in most cases, can fully state them afterwards. The model has found regularities in the examples, and those regularities are the program.

This is the entire conceptual core. Everything else — the architectures, the scale, the terminology — is refinement of that one move: derive the pattern from examples rather than specifying it in advance.

Why the data is the product

Because the pattern comes from the examples, the examples determine what is learned. This is not a caveat; it is the central fact, and most of what goes wrong follows from it.

A model trained on data that under-represents a group will perform worse for that group. A model trained on historical hiring decisions will reproduce the patterns in those decisions, including the ones nobody would endorse if stated openly. The system has no concept of fairness to violate. It has found what was reliably present in what it was shown.

This is why the phrase "the algorithm decided" is usually misleading. The algorithm found a pattern in a dataset that people chose, labelled and assembled. The decisions were made earlier and less visibly.

Correlation is all there is

A model learns which things go together, not why. It has no causal account, and it cannot distinguish a genuine relationship from an accidental one that happened to hold in the training data.

The standard cautionary examples are instructive. A model trained to distinguish wolves from huskies turned out to be detecting snow in the background, because the wolf photographs mostly had snow. A medical imaging model learned to recognise which hospital had taken a scan, because a particular machine's images tended to come from a site treating sicker patients. In both cases the model was highly accurate on the test data and had learned something entirely beside the point.

These are not exotic failures. They are the default risk, and they are why performance on a held-back sample is necessary but not sufficient. If the shortcut is present in the test data too, the test will not catch it.

Generalisation is the whole difficulty

It is easy to build a model that performs perfectly on data it has already seen — it can effectively memorise. That is called overfitting, and it is worthless, because the point is performance on new cases.

The tension is that a model complex enough to capture real structure is also complex enough to memorise noise. Much of the practical craft consists of managing this: holding back data the model never sees during training, deliberately limiting complexity, and stopping before memorisation sets in.

The consequence worth carrying is that performance figures are only meaningful relative to how similar the new cases are to the training data. A model performing at 95 per cent accuracy is making a claim about conditions resembling those it learned from, and that claim can degrade sharply when conditions shift.

What neural networks add

Neural networks are one family of models, loosely inspired by connected neurons — loosely being the operative word. Units arranged in layers pass values forward, each connection carrying an adjustable weight, and training consists of nudging those weights to reduce error.

Their advantage is that with enough layers they can represent very complicated relationships without anyone specifying the form in advance. Earlier approaches often required a human to decide which features mattered; deep networks can derive useful intermediate representations themselves. This is why they took over image, speech and language processing.

The disadvantage is interpretability. A model with billions of parameters offers no readable account of its reasoning, which is a serious problem in any setting where a decision must be explained or contested.

What this does and does not imply

Two opposite errors are common, and both are worth avoiding.

The first is dismissal — that pattern-matching over data is trivial and produces nothing meaningful. This is hard to sustain given the results. Systems built this way have produced genuine advances in protein structure prediction, translation and medical imaging, and the outputs are useful regardless of what is happening internally.

The second is over-attribution — treating fluent output as evidence of understanding. A system trained to produce plausible continuations of text is optimising for plausibility, and plausibility and truth are correlated in the training data without being the same thing. This is precisely why such systems produce confident, well-formed, incorrect statements, and why the errors are hardest to catch in areas where the reader cannot already tell.

The reasonable position sits between: these are powerful statistical tools whose behaviour is grounded in the data they were shown, with real capabilities and predictable failure modes. Knowing that the pattern came from examples tells you where to look when something goes wrong — which is nearly always the examples.