What AI is and is not

The term "artificial intelligence" was coined by John McCarthy for the 1956 Dartmouth Workshop, widely considered the founding event of AI as a field. McCarthy defined it as "the science and engineering of making intelligent machines." That definition is deliberately broad, and its breadth has caused confusion ever since.

In practice, virtually all AI systems in production today fall under a narrower category called narrow AI (also known as weak AI). A narrow AI system performs a specific task, often extremely well, but cannot transfer its ability to other domains. A chess engine cannot write poetry. A spam filter cannot drive a car. Watson could answer quiz questions but could not hold a conversation.

The opposite concept, artificial general intelligence (AGI), refers to a hypothetical system that could perform any intellectual task a human can. AGI does not exist. No timeline for its arrival has scientific consensus. When this course uses the term "AI," it refers to narrow AI unless explicitly stated otherwise.

The key phrase is "tasks commonly associated with intelligent beings." This includes recognising objects in images, translating between languages, generating text, playing games, and making predictions from data. It does not require consciousness, understanding, or general reasoning.

Three terms are frequently used interchangeably in the press and in marketing materials. They are not the same thing. They form a hierarchy:

• Artificial intelligence is the broadest category. It includes any system designed to perform tasks associated with intelligence. This includes rule-based expert systems from the 1980s that contained no learning at all.
• Machine learning (ML) is a subset of AI. ML systems learn patterns from data rather than following explicit rules. A spam filter that learns from labelled examples of spam and non-spam emails is an ML system. Tom Mitchell's 1997 definition remains the standard: a system learns if its performance on some task improves with experience.
• Deep learning (DL) is a subset of ML. It uses neural networks with many layers (hence "deep"). Most recent breakthroughs, including large language models and image generators, are deep learning systems.

This hierarchy matters because different problems require different approaches. Not every problem needs deep learning. Many production systems use simpler ML methods (logistic regression, decision trees, gradient boosting) that are faster to train, easier to explain, and cheaper to run.

Traditional software follows explicit rules written by a programmer. If the input matches condition A, do X. If it matches condition B, do Y. The programmer anticipates every case and writes instructions for each.

Machine learning inverts this. Instead of writing rules, the programmer provides data (examples of inputs and desired outputs) and an algorithm. The algorithm discovers the rules by finding patterns in the data. This is why data quality matters so much, a topic we examine in detail in Module 2.

This inversion has practical consequences:

1. Behaviour is learned, not specified. You cannot read the code to understand why the system made a particular decision. The "logic" is encoded in millions or billions of numerical weights.
2. Performance depends on data, not just code. Better data often improves an ML system more than better algorithms. This is the opposite of traditional software, where performance depends on the quality of the code.
3. Systems degrade over time. The world changes, but a trained model does not update itself. A fraud detection model trained on 2022 patterns may miss 2025 attack vectors. This is called model drift and it requires continuous monitoring.

In 1950, Alan Turing proposed what he called the "imitation game": if a human evaluator cannot reliably distinguish between a machine's responses and a human's, the machine can be said to exhibit intelligent behaviour. This has been popularised as the Turing Test.

The Turing Test is historically important but practically limited. Modern language models can fool humans in short conversations, yet they cannot reliably perform basic arithmetic, maintain consistent beliefs across a conversation, or explain their own reasoning. Passing the Turing Test tells you about surface appearance, not about capability or understanding.

Better evaluation approaches exist and are used in practice. These include task-specific benchmarks (can the system correctly answer medical questions?), adversarial testing (can you find inputs that break the system?), and calibration testing (when the system says it is 90% confident, is it correct 90% of the time?). We cover evaluation methods in depth in Module 5.