Artificial intelligence: programs that learn

A calculator is not AI. It follows rules that a programmer wrote, and it cannot do anything those rules do not cover.

Artificial intelligence (AI) is the branch of computing that builds systems able to perform tasks that normally require human intelligence — recognising a face, understanding speech, translating a sentence, choosing a move in a game, or spotting a tumour in a scan.

The key difference from ordinary software is where the rules come from:

Suppose you want a system that recognises photographs of cats.

1. Collect data. Thousands of photographs, each labelled "cat" or "not cat" by a human. These labels are the ground truth the system learns from.
2. Train. A training algorithm adjusts the model's internal numbers (its parameters) again and again so that its answers match the labels more and more closely.
3. Test. The model is tried on photographs it has never seen. This matters: a model that has memorised the training set can score perfectly there and still fail in real use.
4. Use. The trained model classifies new photographs. Its output is a prediction, with a probability, not a certainty.

Nothing in this process understands what a cat is. The model has simply found patterns in pixel data that correlate with the label.

Expert systems and machine learning are different in an important way: an expert system's rules are written by people and can be read, while a neural network's rules are numbers inside the model and cannot be read directly.

AI is attractive when the rules are too complex or too numerous to write by hand — nobody can write down every rule for recognising a face or understanding a sentence.

Its strengths:

• it finds patterns in large, messy data that people would miss;
• it keeps working at scale, all day, without fatigue;
• it improves as more (good) data becomes available;
• it can handle several inputs at once, such as image plus sound plus text.

Its limits:

• it needs a lot of representative, high-quality data; poor data gives poor predictions;
• it can be biased, because it copies patterns in its training data — including unfair ones;
• it is often a black box: even its developers may not be able to explain a particular decision;
• it is not reliable outside its training range and can fail confidently;
• it costs money and energy to train and run large models.

AI makes old ethical questions bigger, because decisions are made at scale and by systems that cannot explain themselves.

• Bias and fairness. A recruitment model trained on past hiring can repeat past discrimination. The output looks objective, which makes the unfairness harder to challenge.
• Accountability. If an autonomous system harms someone, who is responsible — the developer, the trainer of the data, the operator, or the organisation that deployed it? A professional cannot hide behind "the algorithm decided".
• Privacy. Training data often contains personal information, and models can leak it or be used to identify people.
• Transparency. A decision that affects someone's loan, job or medical care should be explainable to them; a black box makes that difficult.
• Employment. Automation removes some jobs and changes others, with real effects on people's lives.
• Misuse. Deepfakes, automated disinformation and voice cloning are AI used to deceive.