What Is Artificial Intelligence?

The question

Can machines think?

In 1950 the mathematician Alan Turing opened a paper with that question — and immediately set it aside. "Think" and "machine" have no precise definitions, so the argument could go on forever. Instead he proposed a game: an interrogator types questions to two hidden players, a person and a machine. If the interrogator can't reliably tell which is which, on what grounds would we deny the machine intelligence?

The imitation game, now called the Turing test, set the tone for the field that followed: judge intelligence by what a system does. It also contained a warning that is more relevant now than ever — behaving like something that understands is not the same as understanding.

ConceptThe Turing TestUnderstandMust know

The Turing test replaces the vague question 'Can machines think?' with a concrete one: can a machine's conversation be told apart from a human's?

Open the concept page →

At the end of the same paper, Turing suggested that rather than programming an adult mind directly, it might be easier to build a "child machine" and teach it. Hold on to that idea. It took decades to become the dominant approach.

1956

A field is born

In the summer of 1956 a small group of researchers gathered at Dartmouth College for a workshop proposed by John McCarthy, Marvin Minsky, Nathaniel Rochester and Claude Shannon. Their proposal named the field — artificial intelligence — and stated its founding conjecture: that every aspect of learning or intelligence could, in principle, be described so precisely that a machine could simulate it.

What followed was two decades of optimism, and an approach that would dominate AI for thirty years: symbolic AI. Intelligence, the idea went, is the manipulation of symbols — facts, rules, goals — by general procedures like logic and search.

ConceptSymbolic AIKnow wellMust know

Symbolic AI represents knowledge as explicit symbols and rules written by people, and produces intelligent behaviour by manipulating them — through logic and search.

Open the concept page →

How it worked

Intelligence as symbols and rules

A symbolic system knows only what it has been told, in an explicit form: Socrates is a man; all men are mortal. An inference engine applies rules to facts to derive new ones: Socrates is mortal. Every conclusion can be traced back to the rules that produced it — a property modern learned models often lack.

To reason about the world, the system also needs a model of it: categories, properties, relations. Researchers invented semantic networks, frames and ontologies to write this knowledge down — ideas that survive today in knowledge graphs and database schemas.

ConceptLogic and RulesUnderstandMust know

Rule-based AI stores knowledge as if–then rules and derives conclusions by chaining them together, forward from facts or backward from goals.

Open the concept page →

ConceptKnowledge RepresentationUnderstandShould know

Knowledge representation is the problem of writing down what a system knows — objects, categories, relations, defaults — in a form a machine can reason with.

Open the concept page →

Intelligence as search

The other pillar was search. Many problems — a puzzle, a route, a proof, a chess game — can be described as a start state, a set of possible actions and a goal. Solving the problem means finding a sequence of actions that reaches the goal. The trouble is that the number of possible sequences explodes: with 35 options per move, looking ten moves ahead already means about $3 \times 10^{15}$ possibilities.

The answer was the heuristic: a rule of thumb for which possibilities look promising. Try it below — all three algorithms find the goal, but compare how much of the grid each one has to explore.

ConceptSearchKnow wellMust know

Search solves a problem by exploring sequences of possible actions from a start state until one reaches the goal — and a good heuristic decides which possibilities to explore first.

Open the concept page →

Combine search with a model of actions — what each one needs and what it changes — and you get planning. SRI's robot Shakey (1966–72) planned its own routes and actions this way, and the project gave the world the A* algorithm.

ConceptPlanningUnderstandShould know

Planning means finding a sequence of actions that turns the current situation into a goal situation, using a model of what each action requires and changes.

Open the concept page →

1970s–1980s

Expert systems

General problem-solvers turned out to be weak on real problems. By the 1970s a new insight took hold: in practice, expertise comes from detailed domain knowledge. So researchers sat with experts, extracted their know-how as hundreds or thousands of if–then rules, and built expert systems.

They worked — in their niches. Stanford's MYCIN diagnosed blood infections about as acceptably as specialists in a 1979 evaluation. DEC's XCON configured computer orders from 1980, grew to about 2,500 rules, and was estimated to save the company $25 million a year. A commercial boom followed: AI companies, specialized hardware, corporate AI departments.

ConceptExpert SystemsUnderstandMust know

Expert systems captured a human specialist's knowledge as hundreds or thousands of if–then rules, and were AI's first big commercial success.

Open the concept page →

The cold

Winter

Twice, AI's promises ran far ahead of what it could deliver, and twice the money disappeared. The first AI winter (roughly 1974–1980) followed disappointing results in machine translation and critical reviews such as the UK's 1973 Lighthill report. The second began in 1987, when the market for specialized LISP machines collapsed and companies discovered how expensive expert systems were to maintain.

ConceptAI WintersUnderstandMust know

AI winters were periods when inflated expectations collided with limited results, and funding and interest in AI collapsed for years.

Open the concept page →

The problem

Why writing intelligence down failed

The deeper cause wasn't hardware or funding. It was the knowledge-acquisition bottleneck. The hard part of a rule-based system is not reasoning with knowledge; it's getting the knowledge in.

• Experts can't articulate what they know. Recognizing a face, reading an X-ray, understanding a sentence — we do these without being able to state the rules.
• Rules have exceptions, which have exceptions. Birds fly — except penguins, and injured birds, and… Each patch adds rules that interact in unexpected ways, and systems become brittle.
• Common sense is enormous. Everyday reasoning relies on millions of unwritten facts. Encoding them by hand never came close to finishing.

ConceptThe Knowledge-Acquisition BottleneckKnow wellMust know

The knowledge-acquisition bottleneck is the discovery that the hardest part of rule-based AI is getting the knowledge in: much of what experts know is tacit, full of exceptions, and too vast to write down.

Open the concept page →

The idea

From rules to learning

If the rules can't be written, perhaps they can be learned. Give a system many examples with the right answers, a model with adjustable parameters, and an algorithm that tunes those parameters to fit the examples. The knowledge ends up in the parameters, and nobody has to write it down.

Try both approaches on the same problem. First write a spam filter by hand; then watch new messages arrive; then let a learning algorithm build one from examples.

ConceptFrom Rules to LearningKnow wellMust know

Traditional programs turn rules and inputs into answers; machine learning turns inputs and answers into the rules — learning the knowledge from examples instead of having it written in.

Open the concept page →

This wasn't a new idea — Rosenblatt's perceptron (1958) and Samuel's checkers player (1959) both learned. What changed over the following decades was the amount of data, the speed of computers, and better algorithms. By the 1990s statistical machine learning was quietly winning on practical problems; in 2012 deep learning broke through in image recognition; and the story that leads to today's language models began.

Why it matters

Why it matters

It explains why modern AI looks the way it does. Data-hungry models, enormous training runs and the constant search for better datasets are all consequences of choosing to learn knowledge rather than write it.

Symbolic ideas never left. Search, planning, logic and knowledge graphs run inside maps, compilers, databases and games — and today's AI agents are often a learned model wrapped in symbolic scaffolding: tools, plans, structured outputs, verifiers. Whether reasoning in modern models needs more explicit symbolic structure is an open research question Active research.

It is a lesson in reading claims. Both AI winters followed a gap between impressive demonstrations and reliable capability. The habit this site builds — separating established evidence from interpretation and speculation — is the practical lesson of this history.