Lesson 2 — What is Artificial Intelligence?
Most definitions of AI are either so vague they cover a pocket calculator, or so lyrical they describe a film. Here is one that holds up.
A definition you can defend
Artificial Intelligence is the set of techniques that let a machine carry out a task normally requiring human intelligence, by deriving its behaviour from data and experience rather than from instructions written by hand for every case.
Three parts of that sentence are load-bearing.
"A task normally requiring human intelligence." This is deliberately relative, and it has an uncomfortable consequence: the definition of AI moves. Optical character recognition was cutting-edge AI in 1970; today it is a feature in your phone's camera and nobody calls it intelligent. This drift has a name, the AI effect: as soon as a problem is solved, it stops counting as intelligence.
"Deriving its behaviour from data." This is what separates AI from ordinary software. A payroll program applies rules a human wrote. A spam filter derives its own criteria from millions of messages already classified. Same computer, entirely different relationship to knowledge.
"Rather than instructions written by hand for every case." Lesson 1 explained why: for many tasks, nobody can write those instructions.
Narrow AI and general AI
This distinction is the single most useful thing in this lesson, because it lets you tell a serious claim from a marketing one.
Narrow AI — everything that exists
Narrow AI is designed for one task and is helpless outside it. Every system in production in 2026 falls in this category, without exception.
- A model that detects diabetic retinopathy on retinal photographs matches an ophthalmologist. It cannot read a chest X-ray.
- A chess engine crushes every human alive. It cannot play draughts without being rebuilt.
- A large language model writes decent code and passes law exams. It cannot control a robot arm to pick up a cup.
Narrow does not mean weak. It means specialised. Narrow AI already surpasses humans on a growing list of specific tasks, and that is exactly where its commercial value lies.
General AI — what does not exist
Artificial general intelligence would mean a single system that transfers its competence across arbitrary domains as a human does: learning a new task from a handful of examples, reasoning about situations it has never encountered, knowing when it does not know.
It does not exist. No serious researcher can date its arrival, and estimates from credible people span from "within a decade" to "not this century". When a product claims general intelligence, it is selling narrow AI with a broader vocabulary.
Modern language models are extraordinarily convincing generalists in conversation, which makes them feel general. They are not. Their apparent breadth comes from having been trained on an enormous slice of human text, not from an ability to reason across domains. Ask one to handle a genuinely novel situation and the illusion breaks quickly.
Does AI understand what it is doing?
No, and this is worth being precise about because the answer shapes what you can safely delegate to a model.
A model captures statistical regularities in its training data and extends them to new inputs. When a translation model renders "il pleut des cordes" as "it is raining cats and dogs", it has not grasped a metaphor. It has learned that these two sequences occupy equivalent positions in millions of aligned texts.
That is genuinely enough for a great deal. It is not the same as understanding, and the difference surfaces in four recurring ways.
It has no notion of truth. A language model produces plausible text, not true text. When it lacks the information it does not stop; it generates something that has the shape of an answer. This is called a hallucination, and it is not a bug to be patched — it follows from how the model works.
It cannot explain itself. A deep network holds millions of numerical parameters. It can tell you a tumour is present with 94% confidence and cannot tell you which detail convinced it. In medicine, credit or hiring, this opacity is a legal problem as much as a technical one.
It inherits every bias in its data. A recruitment model trained on ten years of a company's decisions will reproduce those decisions, including the discriminatory ones. It is not neutral; it is faithful, which is worse.
It fails silently outside its training range. Shown something genuinely unlike anything it has seen, a model does not report an error. It returns a confident, wrong answer. The AI Ethics course goes into this in depth.
Before delegating a decision to a model, ask: what happens when it is wrong, and will anyone notice? If a mistake is cheap and visible, such as a poor film recommendation, ship it. If a mistake is expensive and invisible, such as a wrongly refused loan, a human must stay in the loop.
Where the word came from
The term was coined in 1956, at a summer workshop in Dartmouth, New Hampshire. The organisers — John McCarthy, Marvin Minsky, Claude Shannon and Nathaniel Rochester — proposed a two-month project for ten people, and wrote in their funding request that a significant part of the problem could be solved in that time.
Seventy years later the problem is not solved. The episode is a useful reminder that in this field, optimism about timelines has a long and consistent track record of being wrong.
In three sentences
Artificial Intelligence is any technique that lets a machine derive its behaviour from data rather than from hand-written rules, which is why the boundary of what counts as AI keeps moving. Everything deployed today is narrow AI: excellent at one task, useless beyond it, with general AI remaining an open research question rather than a product. No model understands what it does, which is acceptable when a mistake is cheap and visible, and dangerous when it is not.