What Is Artificial Intelligence?

No jargon — what artificial intelligence actually is, how ChatGPT-style AI differs from AI that's quietly run things for decades, and what it still can't do.

Artificial intelligence gets used to describe everything from the autocomplete on your phone to a robot uprising in a film — which makes the term nearly useless until someone actually defines it. Here is a definition that holds up: what AI is, why 2022 changed which kind everyone means by it, and what it can and can't actually do.

The short answer

Software that makes a judgement call, not just follows one

Artificial intelligence is software that produces an output a human would normally have to think about, rather than one fully specified in advance by a programmer. A traditional program is a list of exact instructions: if this, then that. An AI system is trained on examples and learns to produce a reasonable answer to situations nobody explicitly wrote a rule for.

That's it. Everything else — the hype, the philosophy, the arguments about whether it 'understands' anything — sits on top of that one idea: a system that generalises from patterns instead of following a fixed rulebook.

Not one thing

AI has been quietly running your life for twenty years

When people say 'AI' today they almost always mean one specific kind — the ChatGPT and Claude kind. But AI as a field is decades old, and most of it is invisible. Your bank's fraud detection, Google Maps picking a route, Netflix's recommendations, spam filters, your phone unlocking at your face — all AI, none of it new, none of it writing essays.

What actually changed in 2022: Generative AI — models that produce new text, images or code rather than just scoring or sorting something — got good enough that the output looked genuinely useful to an ordinary person typing plain English. That's the version everyone means now, and the version the rest of this guide focuses on.

For the numbers behind this: Stanford's AI Index tracks adoption, investment and capability trends across the whole field every year, not just the generative-AI slice everyone talks about. Stanford AI Index Report

Under the hood, briefly

What a model like Claude or ChatGPT is actually doing

A large language model (LLM) is trained on a very large amount of text and learns the statistical patterns of how language fits together — what word, phrase or idea plausibly follows another. When you send it a prompt, it isn't looking anything up or reasoning the way a person does; it's generating the most plausible continuation of your prompt, piece by piece, based on those patterns.

That sounds unimpressive until you notice what 'plausible continuation' means in practice: a plausible continuation of 'here is a summary of this contract' is an actual summary. A plausible continuation of 'here is working Python code that does X' is often working code. There's no trick beyond the model having seen an enormous number of summaries, contracts, and working code.

The part that catches people out: Because the model is generating a plausible answer rather than checking a fact, it can produce something confident, fluent, and wrong — a made-up statistic, a citation that doesn't exist, a function that doesn't run. This is called hallucination, and it's the single most important habit to build around using one: verify anything the output that actually matters.

Want the deeper technical version? IBM's explainer on large language models goes further into training and architecture than this guide needs to. IBM: What are large language models?

Strengths and limits

What it's genuinely good at — and where it lets you down

Where it earns its keep

• Drafting and rewriting text — emails, posts, a first pass at almost anything
• Summarising something long into something short
• Explaining a concept in plain language, at whatever level you ask for
• Brainstorming options you hadn't thought of
• Writing and debugging code, especially the repetitive parts
• Translating and adjusting tone

Where it lets you down

• Being precisely right about facts, dates or numbers it has to recall rather than reason through
• Knowing anything after its training cutoff, unless it's been given search or a document to read
• Exact arithmetic on large numbers, reliably — better with tools, still worth checking
• Telling you when it doesn't know something — it tends to guess fluently rather than say so

Myths

Five things worth unlearning

It's thinking, or conscious

It's pattern completion at a very large scale, not thought. There's no inner experience and no opinion it 'actually' holds between conversations — every reply is generated fresh from your prompt and, at most, the conversation so far.

It knows everything

It knows what was in its training data, up to a fixed cutoff, plus whatever you put in the prompt or it's able to look up. Ask it about your business, your industry, or this morning's news, and it can only help if it can search or you tell it.

It's right whenever it sounds confident

Confidence and correctness are unrelated in an LLM's output. Fluent and wrong is the default failure mode, not the exception — treat confident-sounding output the same whether it happens to be right or wrong, and check anything you'd be embarrassed to get wrong.

It will replace all jobs

It's very good at specific sub-tasks — drafting, summarising, sorting — and much worse at judgement, accountability, and anything that requires actually being responsible for the outcome. What it changes is which parts of a job take five minutes instead of fifty.

It's basically a smarter search engine

Search retrieves existing documents. A language model generates new text that resembles its training data — it can be right, wrong, or a plausible blend of both, and it won't show you sources unless it's specifically built to search and cite.

Read: What is an AI agent? — Once you know what these models actually do, the next question is how businesses wire them up to do real jobs on their own — reading, judging and acting without you opening a tab. That's what an AI agent is.

Frequently asked questions

What is the difference between AI and machine learning?

Machine learning is one method used to build AI — software that improves at a task by learning from examples of data rather than being explicitly programmed. Nearly all AI you interact with today, including large language models, is built using machine learning; 'AI' is the broader goal, 'machine learning' is the technique used to get there.

Is ChatGPT the same thing as AI?

ChatGPT is one product built on one kind of AI — a large language model. AI is the much broader field that also includes recommendation engines, fraud detection, image recognition, and plenty of technologies that have nothing to do with generating text.

Can AI think or feel?

No, not in any sense that matches human experience. A language model generates text by predicting plausible continuations based on patterns in its training data. It has no memory between separate conversations, no goals of its own, and no experience — it responds because you sent a prompt, not because it wanted to.

Why does AI sometimes make things up?

Because it's generating a statistically plausible answer, not retrieving a checked fact. This is called hallucination, and it happens most often with specific details — names, dates, citations, statistics — that the model doesn't actually 'remember' but produces anyway because it sounds like the kind of thing that should follow. Always verify anything factual that matters.

Is AI free to use?

Most major models — Claude, ChatGPT, Gemini — have a free tier that's genuinely usable for everyday tasks, with paid tiers for higher usage limits, more capable models, or business features. For light personal use, you can go a long way without paying anything.