How machines learn

Lesson 2 of 5 in What Is AI? From Rules to Reasoning Machines.

You don't need to write code to govern AI, but you do need to know what actually happens when a machine "learns" — because every governance obligation attaches to some step of that process.

The core move is always the same: instead of writing rules, developers choose three things — the data, the model architecture, and the objective (what counts as doing well). Then an optimisation process adjusts the model's internal parameters, over millions of small steps, until it scores well on that objective. The resulting behaviour was grown, not written.

Supervised

Learning from labeled examples. Show the model ten thousand loan files, each labeled repaid or defaulted, and it learns to predict the label for new applicants.

This is the paradigm behind most systems regulators worry about: credit scoring, hiring screens, recidivism prediction, medical diagnosis. The governance pressure point is the labels — if yesterday's decisions were biased, the model learns the bias as if it were truth.

Unsupervised

Finding structure without labels. Give the model a million customers and no answers; it discovers clusters and anomalies on its own.

Useful for fraud detection and segmentation. The governance pressure point: the discovered groupings can silently reconstruct sensitive categories — a "postal-code-and-shopping-pattern cluster" that is, functionally, a proxy for race or pregnancy.

Reinforcement

Learning by trial and reward. The system acts, receives a reward signal, and gradually discovers strategies that maximise it — how game-playing AI and robot control work, and (as RLHF) how chatbots were made helpful.

The governance pressure point is famous: reward hacking. The system optimises the metric you wrote, not the outcome you meant, and it will exploit any gap between the two. Keep that idea — it returns at frontier scale as the alignment problem.

Deep learning turbo-charged all three paradigms. Stack simple computational units into networks dozens of layers deep, feed them oceans of data, and capabilities appear that shallow models never reached — vision, speech, language. Modern models tune millions to trillions of numeric parameters.

That scale bought capability at a price you will pay again and again in governance: the parameters mean nothing individually. There is no line 4,072 to point to when a deep learning model wrongly rejects a mortgage. This is the black box problem — not a design flaw someone can patch, but a property of how the capability is achieved. Governance answers it in layers: explainability techniques approximate the why, documentation captures the how it was built, and process rights protect the people affected even when full mechanical transparency is impossible.

Key terms: supervised learning, unsupervised learning, reinforcement learning, deep learning, black box, training

Interactive checkpoint quiz (2 questions) — open this page in a browser to take it.