Artificial intelligence is software that performs tasks once thought to require human intelligence: reading text, recognizing objects in photos, understanding spoken language, translating between languages, deciding what to show you next. Most people use AI dozens of times a day without noticing it.

The label covers a broad range of systems, which is part of why the term is so confusing. A spell checker is a form of AI. So is the system that decides which posts appear in your social feed, and so is the one that flags fraudulent credit card purchases. They share a family resemblance but little else.

The central idea is simple. Instead of being told exactly what to do, the machine learns from examples.

For most of computing history, programmers wrote explicit rules. If the temperature reads below freezing, send an alert. The computer was fast and obedient but only as smart as its instructions. That approach breaks down for tasks humans find easy and rulebooks find hard. A programmer cannot write a set of rules that recognizes a cat in any photo, at any angle, in any lighting. A rulebook cannot do it.

Modern AI flips the process. Show the system thousands of labeled examples, and it adjusts its own mathematical settings until it can recognize a cat it has never seen. The system learns the pattern from the data rather than from a programmer's rulebook. That is machine learning, and it is the engine behind almost every product that calls itself AI today.

The pattern-finding can operate at a massive scale. Feed a language model a large share of the written internet, and it learns the statistical shape of human prose. Feed an image model millions of captioned photos, and it learns what a sunset tends to look like. The result is software that can generate new text, new images, and even new video on request.

A key distinction separates the systems in use today from the ones in science fiction. The AI you can use today is narrow. It does one thing. The translator does not drive. The chess engine does not write poetry. The customer service bot does not repair a car. Each system is specialized, trained for a single task, and useless outside it.

Artificial general intelligence, or AGI, is the hypothetical version that matches human flexibility across many domains. One system that can reason, plan, learn, and adapt the way a person can. No one has built it. Despite occasional headlines suggesting it is imminent, AGI remains a research goal, not a product. Whether it will arrive, and when, is one of the most contested arguments in computing.

The gap between narrow AI and general AI explains much of the hype. A company's announcement of an AI breakthrough is almost certainly describing a narrow system. Such a system is impressive inside its lane, but it is not a mind. Holding that distinction keeps both the enthusiasm and the panic in proportion.

Part of the confusion is the word itself. AI covers everything from an email filter to a system with billions of adjustable parameters. Marketers exploit the awe the term carries, so products get called intelligent even when they run on rules a programmer wrote by hand. The useful habit is to demand a description of what the software does.

Knowing what AI is not matters more than knowing what it does. The systems feel intelligent. A chatbot that answers in fluent paragraphs invites you to imagine a conscious entity behind the screen. That impression is an illusion of the interface. The model has no inner experience and no understanding of the words it produces. It is a statistical machine that predicts the next word extremely well. The sentence looks meaningful because human language is full of pattern, and the machine exploits that pattern without comprehending it.

The same is true of image recognition. A system that flags tumors in medical scans has never seen a tumor. It learned a statistical relationship between pixel patterns and labels supplied by human doctors. That distinction matters when the system is wrong, and every system is wrong sometimes. A model trained mostly on one demographic group can fail on another. A language model trained on the internet can absorb the internet's prejudices. The failures are not glitches. They are direct consequences of the data the system learned from.

That is the practical question to ask about any AI: what was it trained on? The training data sets the ceiling of what the system can do and the shape of how it can fail.

The second question is what happens when it is wrong. A recommendation engine that misfires costs you a boring evening. A translation error in a legal document costs more. An error in an autonomous vehicle can cost a life. The consequences scale with the task, and the technology itself does not care. The person and the company deploying the system are responsible for the margin of error.

None of this is to say AI is overhyped or unimportant. The opposite is true. Machine learning has already changed how large parts of the economy operate, and the change is still in its early stages. Sorting and summarizing at machine speed is a shift on the scale of the spreadsheet. The technology will keep improving, cheaply and quietly, until it is unremarkable.

A tool is not a person. The most useful mental model is a very fast pattern matcher with no common sense. It can process more examples than any human could in a lifetime, and it can be wrong in ways no human would be. Both facts are true at the same time.

The next time you read a headline about artificial intelligence, name the thing precisely. Ask what task it performs and what data shaped it, then consider how it can fail. Answer those questions and the term stops being mysterious. AI is a technology you already use and will keep using. It is math, applied at scale to language, images, and decisions.