4.1 What artificial intelligence is, and the four ways to define it

What this is and why it exists

Four definitions of artificial intelligence have been in circulation since the field was named, and they disagree about what success even means. That is not a semantic quibble: which definition somebody holds decides what they build, what they measure, and what they will count as progress. This topic gives you a grid to place any definition in, and the argument for why the field settled on one of the four — the only one you can engineer against.

The vocabulary

• Agent — anything that perceives its environment and acts in it.
• Rational — acting to achieve the best expected outcome, given what is known.
• Performance measure — the external standard by which behaviour is judged.
• Imitation game — the test defining intelligence by indistinguishability in conversation.
• Operational definition — one stated so that you could measure whether it is met.
• Constructive — a definition that tells you how to build the thing.

The mental model

Two questions produce four definitions. Is the goal thinking or acting — an internal process, or observable behaviour? And is the standard human performance or ideal performance — doing what people do, or doing what is best? Cross those and you have the grid, which the Stanford Encyclopedia sets out as systems that think like humans, think rationally, act like humans, and act rationally, calling them "Four Possible Goals for AI".

Each quadrant has a research tradition. Thinking humanly is cognitive modelling: build a system whose internal steps match how people actually reason, and validate it against psychological evidence. Thinking rationally is the logic tradition: encode correct rules of inference and derive true conclusions from true premises. Acting humanly is the imitation tradition, of which the famous test is the headline. Acting rationally is the rational-agent view: do whatever achieves the best expected outcome.

Placing a definition in the grid tells you what its author cares about, and that is the practical use of it. Somebody who says a system is intelligent when nobody can tell it from a person is in the acting-humanly box and has committed to human imitation as the standard. Somebody who says it is intelligent when it derives correct conclusions is in the thinking-rationally box and has committed to logic. Neither is wrong; they are answering different questions, and most arguments about whether something "really counts as AI" are two people in different quadrants.

The imitation test deserves its fame and does not survive as a research goal. Turing's 1950 paper in Mind argued, in the encyclopedia's summary, that the question "Can a machine think?" "should be replaced with the question 'Can a machine be linguistically indistinguishable from a human?'" That was a genuine philosophical advance: it replaced an unanswerable question about inner states with one about observable behaviour.

As an engineering target it fails on two counts. It is not reproducible: the result depends on which judges, how long they have, what they know, and how hard they try, so two runs of the "same" test are not the same experiment and the outcome cannot be compared. And it is not constructive: knowing that a system must pass as human tells you nothing whatever about how to build one. Worse, it actively rewards the wrong things — the reliable strategies for passing involve imitating human failings, deflecting difficult questions, making arithmetic mistakes on purpose, and pretending to be distracted. A test you pass by being worse at things is not a target to design toward.

So the field settled on acting rationally, and the reason is that it is the only quadrant you can both measure and build against. It requires you to state a performance measure — externally, in advance — and then an agent is judged by whether its actions maximise the expected value of that measure. Everything is now operational: the standard is written down, the judgement is about actions rather than inner states, and the definition generalises beyond human competence, which matters for the many tasks where human performance is not the ceiling and for the many where it is not even relevant.

Two clarifications keep the rational-agent view from being misread. It does not mean "logical" — rationality includes acting under uncertainty, on incomplete information, with limited time, where the best expected action may be a guess. And rational is not the same as successful: an agent that made the best available decision and lost to bad luck was still rational, which the next topic develops at length.

Then the pattern that explains much of the public conversation about this subject. As soon as a capability works reliably, it stops being called intelligence and becomes a technique. Playing chess was the standard example of machine intelligence until a machine did it well, at which point it became search. Recognising handwriting, understanding speech, translating text, diagnosing from images — each was cited as the thing machines could never do, and each is now a feature. The category "artificial intelligence" keeps being redefined as whatever has not been achieved yet, which produces two errors worth avoiding. It makes the field look permanently unsuccessful, since every success is reclassified out of it. And it makes each new capability seem to arrive from nowhere, when it usually arrived from thirty years of work on a problem people had stopped calling AI. When you read an assertion about what these systems can or cannot do, ask which quadrant the speaker is in and whether the goalposts have already moved.

What you should now be able to explain or do

State the two questions that generate the four quadrants and place any definition in the grid. Name the research tradition belonging to each quadrant. Explain what Turing's paper replaced and why that was an advance. Give both reasons the imitation test fails as an engineering goal, including what it perversely rewards. Argue why the rational-agent definition was adopted, in terms of measurability and construction. Distinguish rational from logical, and rational from successful. Describe the moving-target pattern and the two errors it produces.

Check yourself