What is artificial intelligence?

ChatGPT is a chatbot built on a large language model — an AI trained to predict likely text, which is what lets it answer questions, draft writing and hold a conversation. You type a prompt and it generates a reply one piece at a time, rather than looking up a stored answer. It's the best-known example of this technology, but not the only one: Claude, Gemini and others work the same broad way.

A search engine finds existing pages and points you to them; a language model instead generates a new answer in its own words, predicting likely text rather than retrieving a document. That's what makes chatbots feel fluent and conversational, but it also means they can state things that are wrong, since they aren't quoting a source by default. Some tools now combine both approaches, letting a model draft its answer from live search results rather than from memory alone.

A language model learns from a fixed collection of text gathered up to a certain point, called its knowledge cutoff, and once training stops it has no awareness of anything that happened afterwards. Having a conversation with it doesn't retrain it, so its knowledge stays frozen at that date no matter how you phrase your questions. Some systems get around this by searching the web or being handed current documents at the moment you ask, which is how they can discuss recent events despite the underlying cutoff.

Treat an AI's answer as a capable draft to check, not as a finished authority. Language models generate text that sounds plausible, but they have no built-in mechanism for checking whether what they say is true, so they can state false things fluently and confidently — a failure mode known as hallucination. Verify anything that actually matters, especially precise facts, figures and recent events, before you rely on it.

There is no credible evidence that today's AI models are conscious or sentient. The mainstream scientific view treats them as sophisticated pattern machines rather than minds, even though what counts as genuine 'understanding' is still a real and open debate among researchers. Consciousness specifically is not something current systems are thought to possess, so dramatic claims to the contrary are worth treating with caution.

A token is a small piece of text — often part of a word rather than a whole one — that an AI language model treats as a single unit when it reads input and generates output. Before a model processes your text, it splits it into these pieces, turns each into a number, and works with the numbers rather than the letters themselves. Common words may become a single token while rarer words, long strings of digits, or unusual spellings are broken into several, so the token count of a passage rarely matches its word count. Because a model's memory limit and its usage cost are both measured in tokens, they are the practical unit for gauging how much text a model can take in at once.

A hallucination is when an AI states something false or fabricated while sounding just as confident as when it is right. It happens because these systems generate text by predicting plausible-sounding sequences of words, not by checking claims against a store of verified facts, so a fluent guess and a correct answer can look identical on the surface. The invented material often reads as convincing detail: a citation that leads nowhere, a quote no one said, a figure with no basis. Because the model gives no signal that it has crossed from knowing into guessing, anything it produces is worth verifying against a trustworthy source before you rely on it.

Training data is the collection of examples a machine learning model learns from, and it shapes almost everything the finished model can do. Each example pairs an input with the pattern or outcome the model is meant to pick up on, so during training the model adjusts itself to fit those examples and, ideally, generalises to new cases it has never seen. Because the model can only learn what its examples contain, gaps, errors, or biases in the training data tend to reappear in the model's behaviour. This makes the choice and quality of training data one of the main things that separates a reliable model from an unreliable one.

Machine learning is a subset of artificial intelligence where systems learn patterns from data to improve their performance on a task without being explicitly programmed with rules for every possible outcome. Instead of following hard-coded instructions, a machine learning model is trained on examples — such as labelled images or past sales figures — and uses statistical techniques to generalise from those examples to new, unseen data. The more relevant data the model receives, the better its predictions or decisions typically become, which is why machine learning powers everything from spam filters to recommendation engines.