What is AI? Definitions, the two pillars, the Turing test

A definition you can defend on an exam

Artificial intelligence is the branch of computer science concerned with the automation of intelligent behavior. Read that carefully: AI is *part of computer science*, so it rests on the same machinery as the rest of the field — data structures for representing knowledge, algorithms for applying it, and languages and programming techniques for building working systems. The definition is deliberately practical, and it has one honest weakness worth stating out loud: *intelligence* itself has no agreed, precise definition. AI therefore advances less by settling what intelligence *is* and more by building programs that do things which, when a human does them, we call intelligent.

The two pillars: representation and search

Almost everything in this course hangs off two foundational concerns. Pin them down now — they are the backbone of the whole syllabus and a recurring exam target.

- Knowledge representation (KR) — the problem of capturing the knowledge a task needs in a form a computer can manipulate. The predicate calculus (Module 1), semantic networks, frames, and neural networks (Module 6) are all *representations*, each with different strengths. KR is not just storing facts; it is encoding them so that reasoning over them is possible. - Search — systematically exploring a space of problem states (successive, alternative stages of solving the problem) to find a solution. A chess player weighing moves and a physician weighing diagnoses are both *searching* among alternatives. Newell and Simon argued search is the essential basis of human problem solving, which is why state-space search (Modules 3–4) gets so much weight.

These two pillars recur literally everywhere: a logic KB plus inference is *representation + search over inferences*; a game tree is *representation of positions + search for the best move*. When a later topic feels unfamiliar, ask "what is being represented, and what is being searched?" — it almost always resolves into these two.

A vocabulary starter pack

Three terms appear from day one:

- A heuristic is a useful but potentially fallible problem-solving strategy — a rule of thumb that usually helps but carries no guarantee. *"Check the appliance is plugged in before assuming it is broken"* and *"castle early to protect the king"* are heuristics. Much of what we call intelligence lives in good heuristics, and heuristic search (Module 4) is a central topic. - An agent is an element of a society that perceives limited aspects of its environment and acts on that environment, directly or by cooperating with other agents. Agents are the modern, situated framing of intelligence. - A knowledge engineer is the AI specialist who works with a domain expert to encode that expert's knowledge into a program that behaves effectively and intelligently — the role behind every expert system.

The Turing test (the imitation game)

Proposed by Alan Turing in 1950, the test measures an allegedly intelligent machine against a human: an interrogator communicates by text only with both a machine and a person in separate rooms, and must decide which is which. If the interrogator cannot reliably tell the machine from the human, the machine is deemed to display intelligence.

Worked example — why the design is clever, in three features. Each design choice removes a way the test could be gamed or derailed:

1. Behavioral, not metaphysical. It gives an *objective* standard based on observable behavior, sidestepping the unanswerable question of what intelligence "really is."
2. No appeal to internal mechanism. It blocks the interrogator from being sidetracked by questions of *how* the machine works or whether it is "conscious" — only the answers count.
3. No bias toward living things. By hiding the contestants behind text, it removes prejudice in favor of biological organisms; only the *content* of the responses matters.

The critique you should be able to give. The most famous objection is Lady Lovelace's Objection: a computer "can only do as it is told" and so can never originate anything truly novel. Experience with expert systems pushes back — diagnostic programs have reached conclusions their designers did not anticipate. A second, structural critique: the test only probes *symbolic, conversational* intelligence — it says nothing about perception or manual dexterity — and it forces machine intelligence into a *human* mold, when a machine might be "intelligent" in genuinely non-human ways (e.g. by exploiting a vast, fast, reliable memory). (Searle's Chinese Room, previewed here and revisited in Module 9, presses the same nerve: passing the test by symbol-shuffling need not mean *understanding*.)