Skip to content

What AI is

25 min Concepts No prerequisites

People have been arguing about what counts as artificial intelligence for seventy years, and they have not stopped. This page gives you the arguments, so that when someone says "that is not real AI" you know which position they are taking.

Why this matters

You are about to spend nine months building these systems. You will be asked what you do, by recruiters, by relatives, by clients. The answer "I work in AI" is close to meaningless right now, because the term covers a spam filter and a research programme aimed at machines that think.

There is a second reason, and it is less obvious. A definition decides things. What gets called AI gets funded, taught, regulated, and built. What falls outside the definition gets ignored. When people argue about the meaning of the word they are usually arguing about something else, which is what should be built and who decides.

The most useful definition

The one that has survived longest comes from Russell and Norvig, whose textbook has trained most of the field since the 1990s. They define AI as the study of agents: systems that perceive an environment and act on it.

The agent loop An agent receives percepts from the world and takes actions that change the world. A goal, or performance measure, defines what counts as success. a goal that defines success Agent decides what to do next The world changes when the agent acts percepts actions

This definition is useful because it is testable. Point it at any system and ask three questions. What does it perceive? What can it do? What is it trying to achieve?

A thermostat perceives temperature, acts by switching a heater, and aims at a set point. By this definition it is an agent, just an extremely simple one. That is not a flaw in the definition. Intelligence is a spectrum, not a club with a door.

Try it on ChatGPT

What does it perceive? Text that a person typed. What does it do? Produce text. What is it trying to achieve? Something set by whoever trained it, not by itself.

So it is an agent, but a limited one. It cannot set its own goals, it waits to be prompted, and in its basic form it only touches the world through words on a screen. This is why people call systems like it weak AI, or a co-pilot. Newer versions that can browse, run code, and call other tools move further along the spectrum, which is exactly why the word agent has come back into fashion.

Four things people mean by intelligence

Underneath the arguing, most definitions of AI land in one of four boxes. They differ on two questions. Is intelligence about thinking or about doing? And is the standard human behaviour, or is it success?

Four aspirations for AI A two by two grid. Rows are thinking and acting. Columns are like humans and rationally. Thinking like humans is cognitive modelling. Thinking rationally is logic and proof. Acting like humans is imitation, including the Turing test. Acting rationally is the rational agent view, where most of the field sits today. like humans rationally THINK ACT Cognitive modelling copy how people reason Newell and Simon, 1960s Logic and proof follow the laws of thought Aristotle to expert systems Imitation be indistinguishable from a person the Turing test, chatbots Rational agents do whatever achieves the goal where most of the field sits now

The bottom right box won, for a practical reason. Acting rationally is the only one of the four you can measure. You can count how often a system reaches its goal. You cannot easily count how human its reasoning felt.

This is worth holding on to, because it explains a recurring fight. When a system succeeds by a method no human would use, some people say that is intelligence and others say it is a trick. Both are being consistent, they are just standing in different boxes.

Seventy years in one screen

A short timeline of AI 1950 Turing proposes the imitation game. 1956 the Dartmouth workshop names the field. 1970s and 1980s expert systems encode human rules. 1997 Deep Blue beats Kasparov by search. 2012 AlexNet makes deep learning the default for vision. 2017 onward, transformers lead to modern language models. 1950 Turing asks whether a machine could imitate a person well enough to fool one 1956 A summer workshop at Dartmouth gives the field its name and its agenda 1970s and 1980s Expert systems encode human knowledge as rules, then hit their ceiling 1997 Deep Blue beats Kasparov at chess, by searching rather than understanding 2012 AlexNet wins an image contest by a wide margin and deep learning takes over 2017 onward The transformer architecture leads to the language models everyone now uses

Two things in that list are worth pausing on.

The field has had winters. Twice, expectations ran far ahead of results, funding collapsed, and researchers avoided the term artificial intelligence on their grant applications. People who lived through those are more careful with predictions than people who did not.

Deep Blue started an argument we are still having. After it won, John McCarthy, the man who coined the term artificial intelligence, complained that it had not advanced the science at all. It played well by brute force, not by thinking. If you only care about results, Deep Blue was a triumph. If you care about understanding intelligence, it was a detour. Every discussion about whether large language models understand anything is a rerun of this argument.

Why this moment happened

Nothing in the last decade required a conceptual breakthrough that the 1980s lacked. Neural networks are from the 1950s. Backpropagation was working in the 1980s. What changed is that three things arrived together.

Why AI advanced now Data, computing power and better algorithms arrived at the same time, which produced the current generation of systems. Data the internet made examples cheap Compute graphics cards made training fast Algorithms architectures that scale with both Systems that learn from examples at scale the same old ideas, finally given enough of everything

Remember this when you read that some new result is a breakthrough. Sometimes it is. Often it is an old idea that finally had enough data and enough hardware.

Where language models sit

A large language model is trained to predict the next piece of text, over and over, across an enormous amount of writing. That is the whole training objective. Everything else it appears to do comes out of doing that well.

That description is accurate and it is also unsatisfying, because the results are hard to square with how simple the objective is. Which is precisely the open question. Does predicting text well require something worth calling understanding, or is fluent output just fluent output? Serious people disagree, and you should be suspicious of anyone on either side who sounds completely certain.

What you can say with confidence is narrower and more useful:

  • It is trained on human writing, so it reflects what is in that writing, including the parts we would rather it did not.
  • It has no way to check its output against the world. It produces text that fits the pattern, whether or not the content is true.
  • On its own it acts only by producing text. Connected to tools, search, and code execution, it acts on much more.

Definitions are decisions

Here is the idea worth taking away from this page.

A definition is never neutral. It is an act of boundary drawing.

When we decide what counts as AI, we are not only describing the world, we are deciding what gets built, what gets funded, what gets taught, and what gets regulated.

Define AI narrowly as pattern recognition at scale and the conversation becomes technical. Define it as systems that make decisions affecting people and the conversation becomes about accountability. Same systems. Different boundary. Different set of questions you are allowed to ask.

You will be inside these systems for the rest of your career. Knowing that the boundary is a choice, made by people, is the difference between following the field and thinking about it.

Questions to sit with

Bring your answers to the next session. There is no key at the back of the book.

  1. A system that reaches a goal by a method no human would use. Is it intelligent, or is it a trick? Does your answer change if the goal is winning chess, diagnosing a patient, or approving a loan?
  2. Which of the four boxes above would you build in, if the choice were yours? What does your choice make easy, and what does it make impossible?
  3. If a definition decides what gets funded and taught, who should get to write it? Researchers, companies, governments, or the people the systems act on?
  4. ChatGPT passes the Turing test on most days. Turing proposed that test as the bar for machine intelligence. Has the bar been cleared, or was it the wrong bar?
  5. What is a task you would refuse to let an AI system do, no matter how accurate it became? Try to say why in one sentence.

How the words fit together

You will hear artificial intelligence, machine learning, deep learning and generative AI used as if they were interchangeable. They are not. Each one sits inside the one before it.

How AI, machine learning, deep learning and generative AI relate Four nested circles. Artificial intelligence is the largest and includes search, planning and rule based systems. Machine learning sits inside it and learns rules from examples. Deep learning sits inside machine learning and uses neural networks. Generative AI is the innermost circle. Artificial intelligence search, planning, rule systems Machine learning learns rules from examples Deep learning neural networks Generative AI new text and images every chatbot is AI, but most AI is not a chatbot

Read it from the outside in.

Artificial intelligence is the whole ambition, and most of it has nothing to do with learning. A chess engine that searches millions of positions, a route planner, a system of handwritten medical rules. All AI, none of it learning anything.

Machine learning is the part where the system works out the rules from examples instead of being given them. This is what your programme is mostly about.

Deep learning is machine learning done with neural networks of many layers. It is what made images, speech and language start working around 2012.

Generative AI is the newest and smallest circle. Models that produce new text, images, audio or code rather than sorting things into categories. Large language models live here.

So ChatGPT is generative AI, which is deep learning, which is machine learning, which is AI. Four true statements about one system. When somebody says they work in AI, this diagram is the question you are really asking.

Next

Now the part you will use every day: how a machine learns anything at all.

How machines learn