What Is AI, Really?

“AI” is stamped on everything from rice cookers to research labs. Here is what the word has meant to the people who coined and study it, and how to tell real AI from marketing.

Walk through any electronics shop in Mid Valley or Lowyat Plaza and count the “AI” labels: AI cameras, AI air-conditioners, AI washing machines, even AI rice cookers. Then open the news and read that AI might cure diseases, take our jobs or outthink us all. Can the same two letters really describe both a washing machine that senses the weight of your laundry and a system that writes essays? Surprisingly, yes — and understanding why is the first step to thinking clearly about AI.

Why the word is slippery

Most words name a thing. “Artificial intelligence” names a goal: making machines do things that would need intelligence if a person did them. And “intelligence” itself has no single agreed definition, even for humans. Is it solving puzzles? Understanding a joke? Finding your way home? Knowing when to keep quiet at a family dinner? Because people disagree about intelligence, they disagree about artificial intelligence too.

The name itself was a choice. In 1955 a young mathematician, John McCarthy, wrote a funding proposal with Marvin Minsky, Nathaniel Rochester and Claude Shannon for a summer workshop at Dartmouth College the following year. In it they coined the term “artificial intelligence” and set out a bold bet:

“…every aspect of learning or any other feature of intelligence can in principle be so precisely described that a machine can be made to simulate it.”McCarthy, Minsky, Rochester and Shannon, Dartmouth proposal (1955)

Decades later McCarthy gave a simpler definition: AI is “the science and engineering of making intelligent machines” (What Is Artificial Intelligence?, 2007). Notice that both are about a field of work, not one particular technology. That is why the label stretches so far. The full story of that 1956 workshop is on Dartmouth 1956.

Key point

AI is best understood as a field — a long project of getting machines to do things we associate with minds — rather than a single invention. Chat assistants, face unlock and the route your Grab driver follows are all products of that field, built with very different techniques.

Four ways to think about AI

The most widely used AI textbook, Stuart Russell and Peter Norvig’s Artificial Intelligence: A Modern Approach, sorts definitions along two questions. First: do we care about how a machine thinks, or only about how it acts? Second: do we measure it against humans, or against an ideal of being rational — doing the best thing given what it knows? Two questions, two answers each, gives four approaches. Tap through them.

The goal: behave so much like a person that you cannot tell the difference.

This is the spirit of Alan Turing’s 1950 imitation game, now called the Turing test (more on Alan Turing): if a judge chatting by text cannot reliably tell machine from human, Turing suggested we should stop arguing about whether it “thinks”. Everyday picture: a customer-service chat on a bank’s website so natural you are not sure whether a human is typing.

This guide’s framing Most AI built today is closest to “act rationally”: engineers mostly care whether a system does the job well, not whether it resembles a human mind. Chat assistants are an interesting twist — they are trained to act humanly in conversation, which is part of why people find them so easy to over-trust.

Narrow AI and general AI

A second useful split is about range. Narrow AI is built for one task or a small family of tasks: recognising faces, recommending videos, translating text, playing chess. It can be superhuman at that task and useless at everything else. The chess program that beats every grandmaster cannot tell you whether it is raining.

Artificial general intelligence, or AGI, describes a hypothetical system that could learn and perform most intellectual tasks a person can, switching flexibly between them — the AI of science fiction. Consensus Every AI system in use today is narrow in at least some important ways, even the impressive ones.

Modern chat assistants blur the line. One system can write a poem, explain photosynthesis, draft an email in Bahasa Melayu and help debug a spreadsheet formula. That is far broader than a chess program. Yet the same system may fail at a simple puzzle a child could solve, invent a fact with total confidence, or be unable to learn from yesterday’s conversation. Broad, but uneven — some researchers call this “jagged”.

Is AGI near, far, or a confused idea?

Debated There is no agreed test for AGI and no agreed definition, so claims that it has arrived or is around the corner are hard to check. Some technology company leaders have said they expect AGI within a few years. Speculative Any date you read is a forecast, not a fact.

Other views
  • Near-term optimists: some leaders of AI companies and some researchers argue that steady scaling of today’s methods will reach human-level ability across most tasks within years, not decades.
  • Sceptics about current methods: researchers such as Gary Marcus and Yann LeCun (in different ways) argue that today’s language models lack key ingredients — reliable reasoning, a model of the physical world, durable memory — and that new ideas are needed.
  • Critics of the concept: others, including Melanie Mitchell, question whether “general intelligence” is one coherent thing at all, and warn that the label can hide what systems actually can and cannot do.
  • Worried researchers: some, including Geoffrey Hinton and Yoshua Bengio, think powerful general systems are plausible soon enough that safety work is urgent.

The AI effect: when AI stops being “AI”

Here is a strange pattern. In the 1960s and 70s, playing good chess was seen as a test of intelligence. Once IBM’s Deep Blue beat world champion Garry Kasparov in 1997, many people shrugged: that was “just search”, not real intelligence. The same happened to reading handwritten postcodes, spam filtering, route planning and speech recognition. Each was a celebrated AI breakthrough; today we just call them features.

This is called the AI effect. The computer scientist Larry Tesler is often quoted as saying “AI is whatever hasn’t been done yet”. Tesler himself said he was misquoted, and gave his version as “Intelligence is whatever machines haven’t done yet.” Either way, the point stands: as soon as something works reliably, we tend to stop calling it AI. The historian Pamela McCorduck described the same pattern in her book Machines Who Think. Lessons from History shows how often the goalposts have moved.

The AI effect cuts both ways. It means AI is already all around you, hidden in ordinary tools (see AI All Around You). It also means marketing teams are tempted to stick “AI” on anything with a sensor and an if-statement. A washing machine that adjusts water to the weight of the load is using a simple automatic rule; whether that deserves the label is, frankly, a matter of taste.

Is this AI? Try it yourself

A practical working test: does the system learn from data or make flexible decisions in situations it was not explicitly programmed for? If it only follows fixed instructions — when the timer reaches zero, ring the bell — most people in the field would call it ordinary automation. Tap each tile, decide for yourself, then read the verdict.

The chat assistant in tile 7 is a large language model, explained in What Is a Large Language Model? Disagree with a verdict? Good — the boundary really is fuzzy. The point is to ask “does it learn or decide?” rather than trusting the label on the box.

Quick check

Who coined the term “artificial intelligence”, and when?

McCarthy, with Minsky, Rochester and Shannon, used it in the 1955 proposal for the 1956 Dartmouth workshop. Turing’s 1950 paper asked “Can machines think?” but did not use the phrase.

A route-finding app that does not think like a human but reliably gets you to your goal best fits which approach?

It acts to achieve a goal well, without imitating human thought or conversation — a rational agent.

Why do people say “it’s not really AI” about chess programs today?

Chess was a flagship AI problem for decades. Once solved, it was reclassified as “just computation”.

Which statement about AGI is most accurate?

AGI remains a contested idea. Predictions vary from a few years to never, and people even disagree about what would count.

// Pause and reflect

  1. Find one product at home or in an advert that is labelled “AI”. Using the “does it learn or decide?” test, would you call it AI?
  2. Which of the four approaches matches what you would want from an AI helper — human-like, or simply good at the job?
  3. Has your answer to “what is AI?” from the welcome page changed? How?

Your notes are saved privately in this browser.

Try this

Ask a family member or colleague, “What do you think AI is?” Listen for which of the four approaches their answer sounds like. Most people describe “acting humanly” — robots and talking computers — which is a useful reminder of how much everyday AI goes unnoticed.

Key terms on this page

Sources & further reading