How AI actually works

This is a high-level view of artificial intelligence: what it is, how it learns, and what actually happens when you type a question into an assistant and it answers. It skips the mathematics and keeps the concepts, since the concepts are what most explanations get wrong or leave out. The idea is not new — the field has a history stretching back to the 1950s — but the systems that made AI a daily tool are recent.

The short version: modern AI does not follow instructions a person wrote out step by step. It learns patterns from very large amounts of data, stores what it learned as billions of adjustable numbers, and uses those numbers to make predictions. Everything below unpacks that sentence.

Three terms, nested inside each other

“Artificial intelligence,” “machine learning,” and “deep learning” are often used interchangeably. They are not the same thing: they are three circles, each sitting inside the last.

All deep learning is machine learning, and all machine learning is AI — but not the reverse. A pocket calculator and a 1980s chess program are both AI in the broad sense, yet neither learns anything. When people in 2026 say “AI,” they almost always mean the innermost circle: deep-learning systems, and especially the generative ones that produce text and images.

Two ways to make a machine behave intelligently

There are broadly two strategies, and the history of the field is largely the story of the second overtaking the first.

Write the rules yourself (symbolic AI). The older approach, dominant from the 1950s into the 1980s, has experts encode knowledge as explicit rules: if this symptom and that test result, then suggest this diagnosis. These “expert systems” work well in narrow, tidy domains where the rules can be written down, and struggle everywhere else, because most real-world skills — recognizing a face, understanding a sentence — resist being spelled out. Nobody can list every rule that separates a cat from a dog in a photo.

Let the machine learn the rules (machine learning). The modern approach flips this around. Instead of writing the rules, you show the system many examples and let it infer the pattern. Show it enough labeled photos and it learns, statistically, what “cat” looks like — without anyone ever defining a cat. The rules still exist, but they are discovered by the machine and stored as numbers, not written by a person in advance.

What a neural network actually is

Most modern AI is built on the neural network, an idea loosely inspired by the brain. “Loosely” matters: it is not a brain and does not work like one in detail, but a mathematical structure that borrows a little vocabulary from biology.

A network is made of simple units, sometimes called neurons. Each takes in numbers, combines them, and passes a number on. What makes one connection count more than another is a weight — a number that dials its influence up or down. One unit does something trivial; wire millions together and the combination can represent extraordinarily complex patterns. The units sit in layers: an input layer that receives the data, one or more hidden layers, and an output layer that gives the result, with information transformed a little at each step. “Deep” learning simply means many hidden layers — depth lets early layers pick up simple features (edges, sounds, letters) and later layers combine them into complex ones (faces, words, meaning).

All the knowledge a network has lives in its weights. The architecture — how the units are wired — is designed by people. The weights are not; they are learned.

Training versus inference: learning versus using

A neural network works in two very different phases, and confusing them causes a lot of misunderstanding.

Training is the learning phase. The network starts with random weights and is fed vast amounts of data. Each time it makes a prediction, its answer is compared with the desired one and every weight is nudged slightly to shrink the error next time. Repeat this billions of times and the weights settle into values that capture real patterns. Training is done once, is enormously compute-intensive, and is where nearly all the cost and energy of building a model goes.

Inference is the using phase. Once training finishes, the weights are frozen; running the model on a new input is inference. Every time an assistant answers a question, that is inference on an already-trained model — far cheaper per use than training, though across hundreds of millions of users it adds up.

One consequence follows directly: a trained model does not keep learning from your conversation unless it is deliberately retrained or given external memory. By default, what it knows was fixed at training time — which is why models have a “knowledge cutoff” and can be unaware of recent events.

How models learn from data

Machine learning is only as good as the data it learns from, and there are three broad ways to learn.

• Supervised learning uses labeled examples — inputs paired with correct answers, such as photos tagged “cat” or emails marked “spam.” The model learns to reproduce the labels, then generalizes to new cases. It is the most common and reliable approach, but it needs someone to produce the labels.
• Unsupervised learning uses data with no labels. The model finds structure on its own — grouping similar items, spotting anomalies, surfacing patterns nobody named in advance. It is useful precisely when you do not know what you are looking for.
• Reinforcement learning has no answer key. An agent acts, receives a reward or a penalty, and adjusts to earn more reward over time, the way a person learns by trial and error. It underlies much game-playing AI and robotics, and a variant helps make assistants more useful.

In every case the data is not a minor detail. A model absorbs the patterns, gaps, and biases of whatever it was trained on. If the data is skewed, the model will be too — which is why the composition of training data is one of the most consequential and scrutinized choices in the field.

Large language models and next-token prediction

The systems behind today’s AI assistants are large language models (LLMs): deep neural networks, usually built on an architecture called the transformer (introduced by researchers in 2017), trained on an enormous amount of text.

Their core skill is simple to state: predict the next token. A token is a chunk of text — often a whole word, sometimes a fragment of one. Given a stretch of text, the model outputs its best guess for what comes next, appends that token, and repeats, one token at a time, until the answer is complete. The reply you read back from an assistant was generated left to right, each token conditioned on everything before it.

This is where parameters come in. A parameter is just a learned weight, and an LLM’s ability to predict well lives entirely in its parameters. Modern large models have billions of them, and the largest run into the trillions, though exact counts for the newest frontier models are often not disclosed. For a sense of scale, a widely cited model from 2020 had 175 billion parameters; leading systems since have grown well beyond that.

One subtle point: LLMs are trained by self-supervision. Nobody hand-labels the text; the “correct answer” is simply the word that actually came next, which the model predicts with the rest hidden. Because the label comes free from the text itself, LLMs can learn from essentially the entire public written record.

A raw next-token predictor is not yet a helpful assistant; left alone it will continue text in unhelpful or unsafe ways as readily as useful ones. Turning one into an assistant takes further steps — instruction tuning on examples of good responses, and reinforcement learning from human feedback, in which people rate answers and the model is adjusted toward the ones they prefer. That is the reinforcement idea from earlier, applied to behaviour rather than to a game.

Why scale changed everything

For most of AI’s history, bigger did not reliably mean better. With language models it did. Researchers found that increasing three things together — training data, computation, and the number of parameters — produced steady, surprisingly predictable gains in performance. These patterns are known as scaling laws.

More striking, at large enough scale some models began handling tasks they were never explicitly trained for: translating, summarizing, writing code, working through multi-step problems. These are sometimes called emergent capabilities, though the term is debated — some argue the jumps only look sharp because of how the abilities are measured. What is not disputed is that the leap in scale from roughly 2018 into the mid-2020s turned language models from a research curiosity into general-purpose tools.

Scale is also why AI became so bound up with hardware and energy: training a frontier model means running large clusters of specialized processors for weeks or months, which is a subject in its own right.

What these systems can and cannot do

Language models are genuinely capable: they draft and edit text, translate, summarize long documents, write and debug code, explain ideas, and hold a coherent conversation. Used well, they are among the most flexible tools computing has produced. But their design creates specific, predictable limitations worth understanding plainly.

Hallucination

Because a model generates the most plausible continuation rather than a verified one, it can state false things with complete fluency — inventing a citation, a date, or a quotation that never existed. This is called hallucination, and it is less an occasional bug than a byproduct of how the system works: it is always predicting likely text, and likely is not the same as true.

Weak grounding

By default a model has no direct link to the world. It was trained on text and, unless connected to live tools, reasons only from patterns in that text: it does not inherently know today’s date, cannot see a web page it was not given, and has no senses. Much recent engineering is about bolting on that missing connection — web search, software tools, and access to private documents at the moment of answering.

Sensitivity to how you ask

The same question, phrased two ways, can produce different answers. Because the model is continuing your text, the wording, order, and framing of a prompt genuinely change the output. This is why “prompting” became a skill worth practicing, and why small rewordings sometimes fix — or break — a response.

Do they actually understand?

This is the honest open question, and it deserves an even-handed answer. On one side, these systems manipulate symbols they have no grounded experience of, hold no goals or awareness, and fail in ways no understanding creature would — so “understanding” overstates the case. On the other, they solve genuinely novel problems, track context across long exchanges, and generalize well beyond what they memorized, which is hard to dismiss as lookup. A reasonable middle view: large models have developed real competence at prediction and pattern-completion that reproduces much of the behaviour of understanding, while whether anything is understood the way a person understands it remains genuinely unsettled.

Frequently asked questions

Related reading

Continue the high-level view: A high-level history of AI, What AI runs on, What AI is used for, Where AI is heading, and How AI decides what to recommend.