What Is Artificial Intelligence?

What is artificial intelligence?

Artificial intelligence is software that does things normally requiring human thought — understanding language, writing, recognising images, translating, answering questions. The difference from ordinary software is where the behaviour comes from: instead of a programmer writing a rule for every case, the system learns patterns from examples and applies them to cases nobody anticipated.

One idea explains almost everything

If you take away one thing, take this: a language model is predicting what comes next, over and over, based on patterns in an enormous amount of text.

That is not a diminishing description — predicting the next piece of text well enough requires having absorbed grammar, facts, reasoning patterns, styles and an extraordinary amount about how the world is described. But it is a mechanism, and once you hold it, the behaviour stops being mysterious. Everything in the rest of this guide follows from it.

In particular, it means the system is producing what a plausible continuation looks like. Usually the plausible continuation is also the true one, because the text it learned from was mostly people saying true things. Sometimes it is not, and nothing in the machinery distinguishes those two cases.

Why it is confident when it is wrong

The single most important practical fact about these systems, and the one that surprises people most.

A person who does not know something usually signals it — they hesitate, hedge, say they are not sure. A language model has no separate mechanism that checks "do I actually know this?" before answering. It produces the most plausible-looking continuation, and a fabricated citation, a made-up statistic and a correct fact are all produced by the same process, in the same fluent tone.

This is what people mean by hallucination, and the word is slightly misleading because it suggests a malfunction. It is the normal operation of the thing, showing up in a case where plausible and true came apart.

What to do about it is simple and non-negotiable: anything that matters gets checked. Names, numbers, dates, quotes, citations, legal and medical and financial claims. The more confident and specific the answer sounds, the more this applies, because fluency is not evidence of anything.

Why it cannot count the letters in a word

A small thing that explains a surprising number of odd failures.

These models do not read letter by letter. Text is chopped into tokens — chunks roughly the size of a common word or a piece of one — and the model works with those chunks, not with the characters inside them. It has no direct view of spelling.

So asking how many times a letter appears in a word, or asking for a word with a specific number of letters, produces guesses. So does arithmetic with long numbers: the digits are chunks being continued by pattern, not quantities being calculated. The same mechanism explains why rhyme and wordplay are hit and miss, and why it is worse at these things in languages that got chopped into more, smaller pieces.

The practical rule: give it a calculator for maths. Modern tools can run actual code or call a real calculator, and where that is available the arithmetic becomes correct because it is no longer being predicted.

Training, inference, and why it does not know today

Two separate stages that people collapse into one.

Training happened once, in advance, over months and at enormous expense. The result is a fixed set of numbers — the model. It is finished before you ever type anything.

Inference is what happens when you use it: your text goes in, a prediction comes out. Nothing is being learned. The model is not changed by your conversation.

Three consequences that answer most beginner questions:

• It has a knowledge cutoff. Events after training are unknown to it, and asking about them produces a plausible guess rather than an admission. Tools that can search the web get round this — but then the answer's quality depends on what it found.
• It does not remember you between conversations, unless the product you are using deliberately stores things and feeds them back in. That is a feature built around the model, not the model learning.
• Correcting it does not teach it. Within a conversation it will take the correction into account. Tomorrow it makes the same mistake.

What it can see: the context window

Each time it answers, the model is shown a bounded amount of text — your message, the conversation so far, any documents you attached, and whatever instructions the product added invisibly. That bundle is the context, and it has a size limit.

This explains the two most common frustrations. In a long conversation, early material eventually falls out of the window or gets compressed, and the model "forgets" what you agreed an hour ago. And in a long document, details in the middle get less reliable attention than material at the beginning or the end.

Two habits follow. Start a fresh conversation when the subject changes rather than continuing an enormous thread — the old context is mostly noise that crowds out what matters. And put the important instruction near your actual question, not twenty messages earlier.

How it works, in order

Most AI today is machine learning: feed a system many examples and it finds patterns. The powerful branch is deep learning, using neural networks — layers of numbers adjusted, repeatedly, until predictions get better. The models behind chat assistants are large language models, trained on very large amounts of text to predict what comes next.

• Data — the examples. Quality matters more than volume past a point, and whatever bias is in the data comes out in the behaviour.
• Training — finding the patterns. Enormous compute, done once.
• Fine-tuning and alignment — a further stage that teaches the model to follow instructions and behave helpfully, which is what makes a raw model into an assistant.
• Inference — you ask, it predicts, you get an answer.

Types of AI

• Generative AI — produces new content: text, images, audio, video. The wave that made all of this mainstream.
• Recognition and classification — spotting faces, filtering spam, flagging an unusual transaction. Older, extremely widespread, and mostly invisible.
• AI agents — systems that do not only answer but act: search, call tools, run code, take steps towards a goal. Genuinely useful and the area where mistakes have consequences beyond a wrong sentence.
• AGI — general intelligence at human level across everything. Does not exist, regardless of headlines, and people who study this disagree sharply about when or whether it will.

What it is actually good and bad at

A practical map, which is more useful than any definition:

• Very good at: rephrasing, summarising, translating, drafting, explaining a concept at a chosen level, generating options, extracting structure from messy text, writing routine code, being a patient tutor.
• Reasonable with supervision: research, analysis, planning, critiquing your work, anything where you can check the result.
• Unreliable at: specific facts, figures and citations; arithmetic without a tool; current events; anything requiring it to know what it does not know.
• Not able to: know your situation beyond what you tell it, take responsibility, or have an opinion it will hold tomorrow.

The pattern: it is strong where the material is in front of it and the job is transformation, and weak where it must supply facts from memory. That single distinction predicts most outcomes, and it is why pasting your document in beats asking about your document.

How image generators differ

Everything above describes language models. Image generators work differently enough that it is worth a paragraph, because it explains their particular failures.

The common approach starts with random noise and repeatedly removes it, step by step, steering each step towards your description — a process called diffusion. The model learned this by being shown enormous numbers of images with captions, so what it is really doing is reconstructing what an image matching those words tends to look like.

That explains the familiar problems. Text inside images comes out malformed, because the model learned what writing looks like rather than what it says. Hands and counting are unreliable, because there is nothing tracking how many fingers there should be — only what hands usually look like. And the same character twice is difficult, because each generation starts from different noise, with no memory of the last one.

The practical consequences are the same shape as for text: use it where a plausible-looking result is what you want, and do not ask it for things that need to be exactly right.

Where the behaviour comes from

Because these systems learn from what people wrote and made, they absorb what was in that material — including the parts nobody would have chosen.

That shows up in ordinary ways rather than dramatic ones: assumptions about who does which job, which names and dialects are treated as standard, whose perspective is the default, which topics are covered thoroughly and which barely at all. Non-English languages are usually represented far less, which is why quality drops off in ways the headline claims do not mention.

Developers work against this — that is part of what the alignment stage is for — and it reduces the problem rather than removing it. For a user, the useful response is not suspicion but specificity: say who the answer is for and what context it is in. Left unspecified, the model fills the gap with whatever was most common in its training data, and that default may not be yours.

You already use it

• Maps and navigation — predicting traffic and choosing routes.
• Streaming recommendations — what to watch next.
• Your phone keyboard — autocomplete, which is the same next-token idea in miniature.
• Your bank — spotting a fraudulent transaction in real time.
• Spam filtering — classification, quietly, on everything you receive.

Worth noticing that most of these predate the current wave and none of them is a chatbot. "AI" is a category, not a product.

What happens to what you type

A fair question and the answer depends on the product rather than on the technology.

Some services use conversations to improve their models; many offer a setting to turn that off, and business tiers usually have it off by default. Either way your text is sent to a company's servers and stored for some period. The practical rules: do not paste anything you would not email to a supplier — client data, credentials, unpublished work, anything covered by a confidentiality agreement — and check the setting before you decide the question does not apply to you.

How to start — today

1. Pick one chat tool — ChatGPT, Claude or Gemini. All have a free tier that is genuinely enough to learn on.
2. Ask it something that actually matters to you this week, not a test question. The difference between a toy and a tool is whether you brought real work.
3. Give it context. Who you are, who it is for, what you have already tried, what to avoid. Most disappointing answers are answers to under-specified questions.
4. Paste your material in rather than describing it. It is far better at working with text in front of it than at recalling things.
5. Argue with it. "That is not right because…" produces a better second answer, and it is how you find out where the edges are.
6. Check anything you will rely on. Every time, not just when it sounds uncertain — because it will not sound uncertain.

A first week that teaches you more than reading does

Understanding this properly comes from use, and a week of deliberate use beats any amount of explanation — including this page. A structure that works:

1. Day one — ask it something you already know the answer to. A topic from your own work. You will immediately see both how good it is and where it goes slightly wrong, which is information you cannot get from a subject you know nothing about.
2. Day two — paste something in. A long email, a document, a set of notes. Ask for a summary, then ask a question about a detail. This is the mode it is strongest in and most people never try it.
3. Day three — make it ask you questions. "Interview me about this until you have enough to draft it." Far better output than describing what you want in one go, and it surfaces things you had not thought to mention.
4. Day four — catch it being wrong on purpose. Ask for sources, then check whether they exist. Ask it to count something. This builds the instinct that keeps you safe later.
5. Day five — give it a real job. Something on your actual list, start to finish, and notice where you had to intervene. That list of interventions is your personal map of what it is for.

By the end of that you will have a working sense of the boundary — which is the thing that separates people who find these tools transformative from people who tried one, got a mediocre answer to a vague question, and concluded it was overrated.

The question everyone actually has

Whether this takes your job is a reasonable thing to wonder and deserves a straight answer rather than reassurance.

What is clearly happening is that tasks are being automated faster than whole jobs. The parts of work that are routine production — drafting, formatting, summarising, translating, boilerplate code — are getting cheap. The parts that involve judgement, responsibility, relationships and knowing which problem matters are not, because the mechanism described on this page does not do those things and it is not obvious that scaling it will.

That reshapes roles rather than deleting them, though "reshaped" can be uncomfortable if the routine production was most of what you were paid for. The honest advice is the boring kind: learn to use it well in your own field, because the competition is not the model, it is the person in your role who is fluent with it. And invest in the parts it cannot do — which are, conveniently, the parts most people found more interesting anyway.

Anyone telling you confidently what the labour market looks like in five years is guessing. The people who study this disagree with each other, and that disagreement is the honest state of the question.

Five things people get wrong

• "It is looking things up." Usually not. Unless it explicitly searched, the answer came from patterns, which is why sources sometimes do not exist.
• "It is learning from our conversation." Not the model. Any memory is a feature built around it.
• "It thinks." It produces text that describes reasoning, and that often helps it reach better answers — but there is no inner deliberation to appeal to, and asking why it said something yields a plausible explanation rather than an account of what happened.
• "Getting better means it will stop being wrong." Accuracy improves; confident wrongness is structural, and checking remains part of using it.
• "It is nearly conscious." Fluency is not evidence of an inner life. It is evidence of a great deal of text.