What is AI, really?

Artificial intelligence is one of the most discussed and least precisely defined terms in modern technology. Depending on who is speaking, it may refer to a spam filter, a self-driving car, a chatbot, or a hypothetical machine mind that does not yet exist. That range is not an accident. "Artificial intelligence" has always been a moving label rather than a fixed technical category, and understanding why is the fastest way to see through the noise.

By the end of this lesson you will have a working definition you can defend in a meeting, and a mental model for telling genuine capability apart from marketing.

A definition you can actually use

Here is a serviceable working definition: an AI system is a computer system that performs tasks we normally associate with human judgment — recognizing, predicting, classifying, generating, or deciding — and that derives its behavior largely from data rather than from rules a person wrote out by hand.

Two parts of that definition carry the weight.

The first is tasks we normally associate with human judgment. This is a social boundary, not a technical one, which is why it keeps moving. Optical character recognition was considered artificial intelligence in the 1970s; today it is a feature in your phone's camera app and nobody calls it AI. This is a pattern sometimes called the AI effect: as soon as a problem is solved reliably, it stops counting as intelligence and starts counting as software.

The second is derives its behavior largely from data. This is the part that genuinely distinguishes the current generation of systems. A traditional program encodes a human expert's rules directly: if the transaction is over a certain amount and the country code is unusual, flag it. A machine learning system is instead shown many examples of flagged and unflagged transactions and works out its own internal criteria. Nobody writes those criteria down; they are discovered.

Narrow intelligence, not general intelligence

Every AI system in commercial use today is narrow. It is built and trained for a bounded set of tasks and has no capability outside that boundary. A model that translates between languages with impressive fluency has no ability to plan your week, and no awareness that weeks exist. Competence in one domain does not transfer the way it does in humans.

This matters because public conversation frequently blends narrow systems with artificial general intelligence — a hypothetical system with broad, flexible, human-comparable capability across arbitrary domains. General intelligence remains a research aspiration. Serious researchers disagree substantially about whether it is achievable, by what route, and on what timescale. Any confident specific prediction you encounter should be treated as opinion, not as an established finding.

The systems you will actually evaluate, purchase, and deploy are narrow ones. Their limits tend to be sharp and somewhat unintuitive: a system can be far more capable than you expect on one task and fail on a variation that looks trivially similar to you.

Three families worth distinguishing

Most of what gets called AI falls into one of three families, and knowing which one you are looking at tells you a great deal about how it will behave.

Rule-based systems encode explicit human-written logic. A tax calculation engine or a clinical decision checklist works this way. These systems are predictable and auditable — you can read the rules — but they are brittle and cannot handle situations their authors did not anticipate.

Machine learning systems learn statistical patterns from labeled or structured examples. Fraud detection, demand forecasting, and medical image triage typically work this way. They handle messy real-world variation far better than rules do, but their reasoning is harder to inspect and their quality depends entirely on the data they were trained on.

Generative systems — the large language and image models that dominate current attention — learn patterns from very large collections of text or images and produce new content in response to a prompt. They are strikingly general within the medium they were trained on, and they produce fluent output whether or not that output is correct, which is a failure mode we will treat carefully later in the course.

Many real products combine all three. A customer support tool might use a rule to route billing questions, a classifier to detect urgency, and a language model to draft the reply.

Why the distinction pays off

Knowing which family a system belongs to lets you ask the right questions. For a rule-based system, ask who maintains the rules. For a machine learning system, ask what data it learned from and whether that data resembles the situation you will use it in. For a generative system, ask how you will verify the output before anyone acts on it.

You do not need mathematics to ask any of these questions well. You need a clear picture of what the system is doing and where its behavior came from — which is exactly what the rest of this course builds.