AI basics

What counts as AI?

There is no single, universally agreed definition of intelligence, so practical definitions of AI describe what a system does rather than what it is. The US National Institute of Standards and Technology (NIST) describes an AI system as an engineered or machine-based system that, for a given set of objectives, generates outputs such as predictions, recommendations or decisions that influence real or virtual environments. A spam filter, a route planner, a chess program and a chatbot all fit that description, even though they work in very different ways.

• Objectives: an AI system is built to pursue goals that people set, such as 'flag spam' or 'win the game'.
• Inputs: it takes in data — text, images, sensor readings, game positions.
• Outputs: it produces predictions, recommendations or decisions.
• Influence: those outputs change something, from which email you see to which move is played.

Where the idea came from

In 1950 Alan Turing asked 'Can machines think?'. Rather than define 'machine' and 'think' by how people use the words (which, he wrote, would reduce the question to an opinion poll), he proposed replacing it with a closely related question framed as a game he called the imitation game. Five years later, a proposal dated 31 August 1955 by John McCarthy, Marvin Minsky, Nathaniel Rochester and Claude Shannon called for a two-month study of 'artificial intelligence' at Dartmouth College in the summer of 1956. That proposal was openly built on a conjecture: that every aspect of learning or intelligence could in principle be described precisely enough for a machine to simulate it.

Two ways to build an AI system

• Hand-written rules (the knowledge-base approach): people write the knowledge down as formal rules. Good for problems that can be stated as a short list of exact rules; breaks on situations nobody anticipated.
• Learning from data (machine learning): the system extracts patterns from examples. Good for problems that are easy for people but hard to describe, such as recognising faces or speech; inherits gaps and biases in its training data.

Early projects that tried to hard-code knowledge about the world as formal rules ran into trouble with ordinary situations that nobody had written a rule for. That difficulty is one reason much of modern AI uses machine learning: instead of being told the rules, the system finds patterns in data.

Worked example: is this a rule or a learned pattern?

Imagine sorting emails into 'spam' and 'not spam'. A rule-based filter might say: 'if the subject contains the word FREE in capitals, mark it as spam'. A learning-based filter is shown thousands of emails that people have already labelled and adjusts its own internal numbers until it sorts those examples well. The first is easy to read but easy to fool; the second can catch patterns no one thought to write down, but it is only as good as the examples it learned from.

What AI can and cannot do (as of 2026)

Capability claims go out of date quickly, so they should always carry a date and a source. The Stanford AI Index 2026 reported that an AI system earned a gold-medal score at the 2025 International Mathematical Olympiad, while the best model tested read analogue clocks correctly only about half the time. Researchers call this uneven pattern a 'jagged frontier': strength on one task does not predict strength on a neighbouring one. A year earlier, the AI Index 2025 had flagged that models still struggled on complex reasoning benchmarks such as PlanBench, often failing logic tasks even when provably correct solutions exist.

Using AI responsibly

Because AI outputs influence real decisions, NIST's AI Risk Management Framework sets out what makes an AI system trustworthy: valid and reliable, safe, secure and resilient, accountable and transparent, explainable and interpretable, privacy-enhanced, and fair with harmful bias managed. These qualities can pull against each other, so they have to be balanced for each use.