What Is Artificial Intelligence?

We call ourselves Homo sapiens — man the wise — because our intelligence is what we prize most. For thousands of years philosophy has asked how a mere handful of matter can perceive, understand, predict, and act on a world far larger than itself. Artificial intelligence asks the same question and then goes one step further: not just to understand intelligence, but to build it.¹

That leaves a problem. Build intelligence is a goal only once we fix what counts as success, and the field has never fully agreed. In place of consensus it has a small grid of definitions — eight, sharpened over half a century — that disagree in a structured, informative way. Read the grid first: which cell you stand in decides what you are building and how you would know you had built it.

Four schools along two axes

Every serious definition of AI answers two questions. First, is the target a thought process or a behavior — what goes on inside, or only what comes out? Second, is success measured against human performance or against an ideal standard, rationality (doing the right thing given what one knows)? The two binary axes give four schools.²

Each cell is not a single slogan but a cluster of published definitions — AIMA collects eight, two per cell, drawn from the field's textbooks. They disagree on emphasis in a way that maps cleanly onto the axes: the top row talks about thought and reasoning, the bottom about behavior and action; the left column measures against people, the right against a formal standard.

The distinction between humanly and rationally is not a claim that humans are irrational. It is only the recognition that human performance is a specific, sometimes flawed, empirical target, while rationality is an ideal one can define mathematically. The two lead to genuinely different research programs, and both have been pursued. A human-centered approach is in part an empirical science — observing and hypothesizing about people. A rationalist approach combines mathematics and engineering. We take the four cells in turn.

Acting humanly: the Turing test

Alan Turing (1950) proposed to sidestep the unanswerable question can machines think? with an operational test. A human interrogator holds a written conversation with an unseen partner; if the interrogator cannot reliably tell whether the partner is a person or a machine, the machine passes. The Turing test replaces a metaphysical question with a behavioral one you can actually run.³

Passing it is not a parlor trick. To sustain the conversation a machine would need four capabilities, and these four map almost exactly onto the major subfields of AI:

Turing deliberately kept the interrogator and machine in separate rooms, because physically imitating a person is beside the point. The total Turing test adds a video channel and a hatch through which objects can be passed, forcing two more capabilities: computer vision to perceive, and robotics to manipulate and move. Six disciplines, and a test that has stayed relevant for over sixty years.

Yet the field has spent almost no effort on passing the test. The reason is a recurring theme of the whole subject: the goal is to understand the principles of intelligence, not to counterfeit one instance of it. Aeronautics did not progress by building machines that fly so exactly like pigeons that they can fool other pigeons; it progressed once the Wright brothers stopped imitating birds and started studying aerodynamics.³

The test also invites a family of objections that clarify what it does and does not certify. Turing himself catalogued nine anticipated rebuttals in the 1950 paper and answered each, and two of them still frame the debate. The argument from disability — a machine can never do $X$ for some cherished human $X$ (be kind, fall in love, tell right from wrong) — Turing dismissed as an inductive overreach from no machine has yet to no machine can ever, a bet the last seventy years have repeatedly lost. The deeper objection came later, from Searle's Chinese Room (1980): imagine a person who speaks no Chinese sealed in a room, hand-executing a rulebook that maps Chinese questions to Chinese answers. The room as a whole passes a Chinese Turing test, yet nobody inside understands Chinese — so, Searle argues, passing the test shows behavior, not understanding. Whether that matters depends on which cell of the grid you stand in. For the acting-rationally school it does not: an agent is judged by what it does, and a system that reliably does the right thing is the object of study whether or not there is understanding behind the behavior. That split traces the same thinking-versus-acting axis, now resurfacing as a philosophical dispute.⁴

A cautionary counterweight arrived early. Weizenbaum's ELIZA (1966) was a few hundred lines that impersonated a psychotherapist by reflecting the user's own sentences back as questions (I am unhappy → Why are you unhappy?), with no model of language or mind at all. Users nonetheless confided in it and insisted it understood them — the ELIZA effect, our standing tendency to read intention into surface behavior. It is the reason a conversational imitation, on its own, is weak evidence of intelligence, and the reason the field measures agents against tasks with an external performance measure rather than against a human judge's impression.

Thinking humanly: cognitive modelling

To claim a program thinks like a human, you first need a theory of how humans think — and that theory comes from getting inside the actual mind, through introspection, controlled psychological experiment, or brain imaging. Once the theory is precise enough to express as a program, you can check whether the program's step-by-step behavior matches a person's on the same task.

Newell and Simon's General Problem Solver (1961) was built with exactly this ambition. They were not content to have it solve puzzles correctly; they compared the trace of its reasoning steps against the traces of human subjects. That comparison, of computational models against experimental data, is the whole project of cognitive science. It is a distinct claim from the program performs well, and separating the two — a good engineering result versus a good model of the human mind — is what let AI and cognitive science each advance faster.⁵

Thinking rationally: the laws of thought

Aristotle tried to codify right thinking as irrefutable reasoning: his syllogisms gave argument patterns that always yield correct conclusions from correct premises. Socrates is a man; all men are mortal; therefore Socrates is mortal. The study of such patterns became logic, and the logicist tradition in AI hopes to build intelligent systems on top of it: express the problem in formal notation, run a logical procedure, read off the answer.⁶

A worked instance shows the pattern the whole field inherits. From all men are mortal and Socrates is a man, one modus-ponens step yields Socrates is mortal. Feed a logical engine those two premises and the conclusion falls out immediately. But that is the trivial case: two premises, one relevant rule. The engine's real task is to find the right inference among the many it could draw, and here the count turns against it. With $k$ facts in the knowledge base, the number of derivation chains grows on the order of $O(k!)$; most are irrelevant to the query.

Two obstacles bar the road, and both surface again throughout the course. First, informal knowledge is hard to state in the crisp terms logic demands, especially when it is less than fully certain — birds fly is true until you meet a penguin, an ostrich, or a bird with a broken wing, and no finite list of premises can enumerate every exception. Second — and this is the deeper one — being able to solve a problem in principle is not the same as solving it in practice. A logical system with only a few hundred facts can exhaust any computer's resources unless it has guidance about which reasoning steps to try first. The general problem of deciding whether a set of logical sentences is even satisfiable is NP-complete: as the number of variables grows, the worst-case work grows faster than any polynomial, so just search the possibilities is not an option past a modest size. Correct inference is not enough; tractable inference is the real target, and steering the search — which the informed-search and inference lessons take up — is most of the work.

Acting rationally: the rational agent

An agent is something that acts — the word comes from the Latin agere, to do. Every program acts, but a rational agent is expected to do more: operate autonomously, perceive its environment, persist over time, adapt to change, and create and pursue goals. A rational agent acts so as to achieve the best outcome or, when there is uncertainty, the best expected outcome.⁷

Correct inference — the target of the laws of thought school — is one way to act rationally: reason your way to the conclusion that an action achieves your goals, then take it. But it is not the only way, and sometimes not the best way. In some situations there is no provably correct thing to do, yet something must still be done. And some rational acting bypasses inference entirely: recoiling from a hot stove is a reflex that usually beats a slow, deliberated response.

This is why Russell & Norvig — and this course — adopt the acting-rationally view.

Stated as a choice rule, with performance measure (utility) $U$ over outcome states $s$ and action-conditional outcome distribution $P(s\mid a)$, the rational action is

the action of maximum expected utility. Every mechanism in the course is a way of computing this argmax under some new difficulty.

Two advantages settle the choice. First, rationality is more general than the laws-of-thought approach: correct inference is just one of several mechanisms for achieving rationality, so the rational-agent frame contains the logicist one rather than competing with it. Second, it is more amenable to scientific development. The standard of rationality is mathematically well defined and completely general; you can turn it into agent designs that provably achieve it. Human behavior, by contrast, is adapted to one specific environment and is defined only by the sum total of what humans happen to do.⁷

One honest caveat, which the course returns to repeatedly. Perfect rationality — always doing exactly the right thing — is infeasible in complicated environments, because the computation costs too much. For most of the course we adopt the working hypothesis that perfect rationality is a good starting point for analysis, and treat limited rationality — acting appropriately when there is not enough time to compute the ideal action — as the refinement.

The idea has a lineage. Herbert Simon, one of the field's founders, argued in the 1950s that real agents do not maximize; they satisfice — search only until they find an option good enough against an aspiration level, then stop. He called the resulting account bounded rationality, and it reframes the design goal. The target is not the action a cost-free oracle would choose but the action a real computation, itself consuming time and energy, should choose. Formalizing that gives bounded optimality: an agent program is bounded-optimal if it is the best possible program for its machine, accounting for the cost of the deliberation itself. This is the honest standard the course keeps in the background — perfect rationality frames the analysis, bounded optimality is what any deployed agent can actually aim at.⁸

The notation for the course

The rational-agent frame fixes a small vocabulary used in every later lesson. The whole subject is built out of these terms.

Keep the last row in view. Everything downstream — search, logic, probability, learning — is a different answer to the same question: given the percepts so far, which action maximizes expected performance? The lesson on intelligent agents makes the agent and its environment precise; the rest of the course fills in the machinery an agent uses to choose well.

That is the definition. But a definition arrived without its history, and without the older sciences it borrows from, is hard to trust. The companion lesson, the foundations of AI, traces where the rational-agent idea came from — the disciplines that fed it, the cycles of boom and winter the field went through, and the deep-learning turn that now defines its public face — and shows that none of it disturbs the one-line frame fixed here.

This continues in The Foundations of AI.